Incentives and two-sided matching - engineering coordination mechanisms for social clouds

Haas, Christian

2014

KIT Scientific Publishing

The Social Cloud framework leverages existing relationships between members of a social network for the exchange of resources. This thesis focuses on the design of coordination mechanisms to address two challenges in this scenario. In the first part, user participation incentives are studied. In the second part, heuristics for two-sided matching-based resource allocation are designed and evaluated.

Business

Management

English

9783731502371

Studies on eOrganisation and Market Engineering 12

Christian Haas

Incentives and Two-Sided Matching
Engineering Coordination Mechanisms for
Social Clouds

Christian Haas
Incentives and Two-Sided Matching
Engineering Coordination Mechanisms for Social Clouds

Studies on eOrganisation and Market Engineering
Karlsruher Institut für Technologie (KIT)
Herausgeber:
Prof. Dr. Christof Weinhardt
Prof. Dr. Thomas Dreier
Prof. Dr. Rudi Studer

12

Incentives and Two-Sided Matching
Engineering Coordination Mechanisms for Social Clouds

by
Christian Haas

Dissertation, Karlsruher Institut für Technologie (KIT)
Fakultät für Wirtschaftswissenschaften, 2014
Referenten: Prof. Dr. Christof Weinhardt, Prof. Dr. Steven O. Kimbrough

Impressum

Karlsruher Institut für Technologie (KIT)
KIT Scientific Publishing
Straße am Forum 2
D-76131 Karlsruhe
KIT Scientific Publishing is a registered trademark of Karlsruhe
Institute of Technology. Reprint using the book cover is not allowed.
www.ksp.kit.edu

This document – excluding the cover – is licensed under the
Creative Commons Attribution-Share Alike 3.0 DE License
(CC BY-SA 3.0 DE): http://creativecommons.org/licenses/by-sa/3.0/de/
The cover page is licensed under the Creative Commons
Attribution-No Derivatives 3.0 DE License (CC BY-ND 3.0 DE):
http://creativecommons.org/licenses/by-nd/3.0/de/

Print on Demand 2014
ISSN 1862-8893
ISBN 978-3-7315-0237-1
DOI 10.5445/KSP/1000041861

Incentives and Two-Sided Matching
Engineering Coordination Mechanisms for Social
Clouds

Zur Erlangung des akademischen Grades eines
Doktors der Wirtschaftswissenschaften
(Dr. rer. pol.)
von der Fakultät für
Wirtschaftswissenschaften
am Karlsruher Institut für Technologie (KIT)

genehmigte
D ISSERTATION

von
Dipl.-Wi.-Ing. Christian Haas

Tag der mündlichen Prüfung: 17. Juni 2014
Referent: Prof. Dr. Christof Weinhardt
Korreferent: Prof. Dr. Steven O. Kimbrough

Karlsruhe, 2014

Abstract
An unprecedented variety of resources are shared, exchanged and traded within
(social) networks of users. In this context, the Social Cloud framework leverages
existing relationships between members of a social network for the exchange of
resources. The design of a Social Cloud comprises several challenges that need to
be addressed to create a sustainable platform. This ranges from the identification
and provisioning of relevant incentives for user participation to achieve a critical
mass of users, the understanding and modeling of the underlying trust concepts,
the design of market mechanisms for resource allocation, to implementation details
that ensure the technical feasibility of the platform.
This thesis focuses on the design of coordination mechanisms to address two
of these challenges, namely user participation incentives and resource allocation
mechanisms. The thesis applies a simulation-based approach to design incentive
schemes and allocation mechanisms. In the first part, based on the survey-based
identification of relevant participation incentives and their dependency on certain
factors, two case studies show the usefulness of applying simulations in the engineering of contribution schemes for Social Clouds. The second part of the thesis
advocates the use of two-sided matching for resource allocation due to the social
setting of the considered scenario. For this preference-based allocation, heuristics
are proposed and evaluated for one-to-one matching as a means to provide flexibility with respect to diverse user preferences and objective functions, and to find
high quality solutions in settings for which no efficient exact or approximation algorithms exist.
The case studies and results of this thesis for user participation exemplify that a
simulation-based approach can be leveraged as complementary methodology to
analytical modeling and prototyping to find useful results with respect to the effects of contribution schemes on different user types. Considering resource allocation, the proposed algorithms provide an increased solution quality compared to
existing algorithms, and have the advantage of inherent flexibility with respect to

i

ii
changing requirements and matching goals as they can be easily adapted to different preference or goal settings. The study of preference manipulation shows
that while there are cases where manipulation is beneficial, the expected reward
is comparably small, which indicates little space for practical manipulation in real
scenarios. Furthermore, the thesis considers a dynamic allocation scenario showing
that continuous reallocation is necessary if the amount of supplied and demanded
resources fluctuates.
This work contributes to the field of incentive engineering by providing new insights in relevant incentives for social resource sharing and through the application of simulation-based approaches to design contribution schemes. The field of
two-sided matching is advanced by showing that the proposed heuristics provide
superior performance and flexibility for preference-based resource allocation.

Acknowledgements
The completion of this work was enabled by the support of many people, to whom
I own my deepest gratitude. First of all, I want to thank my supervisor Prof. Dr.
Christof Weinhardt for his ongoing support, advice, and encouragement throughout the PhD at the Institute of Information Systems and Management (IISM) as
well as the Karlsruhe Service Research Institute (KSRI). Furthermore, I am very
thankful to my co-advisor Prof. Dr. Steven O. Kimbrough (University of Pennsylvania, USA) for the insightful discussions and helpful comments that guided this
research. I also want to particularly thank Prof. Dr. Thomas Setzer and Prof. Dr.
Kay Mitusch for serving on my thesis committee.
My sincere thanks go to all my former and current colleagues from IISM and KSRI,
which have been a fundamental factor that made the thesis such an interesting and
enriching endeavor. In particular, my special thanks go to Dr. Simon Caton for
his friendship and the numerous professional and personal meetings that we had
over the years. I also want to thank Margeret Hall, Alexander Schuller, and Anuja
Hariharan for reading and improving parts of the thesis, and who together with
Anders Dalén, Axel Kieninger, and Tim Straub made the PhD time much more
enjoyable through various social events.
For the ability to pursue a research stay at the University of Chicago and Cardiff
University as well as having very fruitful collaborations I owe my honest gratitude
to Dr. Kyle Chard and Prof. Dr. Omer Rana. I also gratefully acknowledge the
financial support by the Karlsruhe House of Young Scientists (KHYS) that enabled
this research stay.
Finally, I thank my family, Hermann, Elisabeth and Michael Haas, for their continuing support and love throughout my entire life. I also thank Magie for her
emotional support, love and constant encouragement. Without them, this work
would not have been possible.

Contents
List of Figures

ix

List of Tables

xi

List of Abbreviations

xiii

List of Symbols

I.

xv

Introduction and Foundations

1. Introduction
1.1. Design Challenges . .
1.2. Research Outline . . .
1.3. Structure of the Thesis
1.4. Research Development

1
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

3
6
11
17
19

2. Resource Sharing in Social Contexts
2.1. The Concept of Social Clouds . . . . . . . .
2.1.1. Definitions . . . . . . . . . . . . . . .
2.1.2. Related Concepts . . . . . . . . . . .
2.1.3. Design Challenges . . . . . . . . . .
2.1.4. A Prototype Social Compute Cloud
2.2. Simulating Social Clouds . . . . . . . . . . .
2.2.1. Purpose and Potential Applications
2.2.2. Requirements and Related Tools . .
2.2.3. Architecture . . . . . . . . . . . . . .
2.3. Coordination Challenges in Social Clouds .
2.3.1. Participation Incentives . . . . . . .
2.3.2. Resource Allocation . . . . . . . . .
2.4. Summary . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

21
22
22
27
28
30
31
32
34
38
40
40
47
51

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

v

vi

Contents

II. Incentive Engineering for Social Clouds
3. Incentives in Social Clouds
3.1. Participation Incentives in Resource Sharing Platforms . .
3.1.1. Incentives in Online Communities . . . . . . . . . .
3.1.2. Participation Incentives in Similar Sharing Systems
3.1.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . .
3.2. Engineering Incentives for Social Clouds . . . . . . . . . .
3.2.1. Incentives During the Participation Lifecycle . . . .
3.2.2. Factors Influencing the Participation Incentives . .
3.2.3. Design Implications for Social Clouds . . . . . . . .
3.3. Identifying Relevant User Incentives . . . . . . . . . . . . .
3.3.1. Goals and Design of Web Survey . . . . . . . . . . .
3.3.2. Evaluation . . . . . . . . . . . . . . . . . . . . . . . .
3.3.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . .
3.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

53
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

55
57
58
60
62
63
63
68
71
72
72
74
81
83

4. Designing Incentive Schemes and Co-operative Infrastructures
4.1. User Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2. Case Study: Designing Incentive Schemes . . . . . . . . . . . . . . .
4.2.1. Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2. Modeling an Incentive Scheme . . . . . . . . . . . . . . . . .
4.2.3. Evaluation of Dynamic Effects . . . . . . . . . . . . . . . . .
4.2.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3. Case Study: Co-operative Infrastructures . . . . . . . . . . . . . . .
4.3.1. Definition of Co-operative Infrastructures and Related Work
4.3.2. Economic Model . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.3. Evaluation of Contribution Schemes . . . . . . . . . . . . . .
4.3.4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.

85
86
90
91
92
97
103
103
105
108
115
122
125

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

III. Two-Sided Matching in Social Clouds
5. Resource Allocation in Social Clouds
5.1. Preference-based Resource Matching . . . . . . . . . . . . . . .
5.1.1. Definitions and Relevant Theorems . . . . . . . . . . .
5.1.2. Performance Metrics for Two-Sided Matching . . . . .
5.1.3. Preference Structures and Computational Complexity .
5.2. Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3. Algorithms for Preference-based Matching . . . . . . . . . . .
5.3.1. Exact Algorithms . . . . . . . . . . . . . . . . . . . . . .
5.3.2. Approximation Algorithms . . . . . . . . . . . . . . . .
5.3.3. Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . .

129
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

131
133
133
136
138
140
144
145
149
151

Contents
5.4. Performance of Matching Algorithms . . . . . . .
5.4.1. Simulation Specifics . . . . . . . . . . . . .
5.4.2. Algorithm Runtime . . . . . . . . . . . . . .
5.4.3. Complete Preferences With Indifferences .
5.4.4. Incomplete Preferences With Indifferences
5.5. Summary . . . . . . . . . . . . . . . . . . . . . . . .

vii
.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

157
158
160
163
168
178

6. Incentive Compatibility and Dynamic Allocation
6.1. Strategic Manipulation in Resource Allocation . . . . . . . . . . .
6.1.1. Theoretical Results and Manipulation Strategies . . . . . .
6.1.2. Effects of Preference Manipulation on Matching Outcome
6.1.3. Robustness of Matching Algorithms against Manipulation
6.1.4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2. Dynamic Resource Allocation in a Social Compute Cloud . . . . .
6.2.1. Approaches to Capture Intermediate Supply and Demand
6.2.2. Evaluating the Effects of Dynamic Allocation . . . . . . . .
6.2.3. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

181
183
183
189
194
200
201
202
203
207
207

IV. Finale
7. Conclusion
7.1. Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1.1. Simulation-based Approach to Study Social Clouds
7.1.2. Understanding Incentives for Participation . . . . .
7.1.3. Heuristics for Preference-based Resource Allocation
7.2. Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2.1. User Participation Incentives . . . . . . . . . . . . .
7.2.2. Preference-based Resource Allocation . . . . . . . .

V. Appendix

211
.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

213
214
214
215
218
222
222
225

231

A. Additional Material for Incentive Survey

233

B. Extended Result Tables for Co-operative Infrastructures

243

C. Extended Results for Preference-based Matching

247

D. Extended Results for Preference Manipulation

257

References

261

List of Figures
1.1. Structure of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

2.1.
2.2.
2.3.
2.4.

Example of a Social Cloud . . . . . . . . . . . . . . . . . . . . . . . .
General Architecture of a Social Cloud . . . . . . . . . . . . . . . . .
Architecture of a Social Compute Cloud and its Core Components .
Layered Structure of the Social Exchange Simulator . . . . . . . . .

.
.
.
.

24
25
31
38

3.1.
3.2.
3.3.
3.4.
3.5.
3.6.

Participation Stages and User Incentivization Problems . . . . . . .
Example of Relationship Types and Resource Sharing Incentives . .
User Groups with which Resources were Previously Shared Online
Incentives for Previous Online Resource Sharing . . . . . . . . . . .
Survey Results for Registration and Participation Incentives . . . .
Survey Results for Ten Item Personality Inventory (TIPI) . . . . . .

.
.
.
.
.
.

64
69
75
76
77
80

4.1.
4.2.
4.3.
4.4.
4.5.
4.6.
4.7.
4.8.
4.9.

Average User Utility . . . . . . . . . . . . . . . . . . . . . . . . . .
Average Offers and Matches per Offer . . . . . . . . . . . . . . . .
Relative Change in User Utility through Trading Constraint . . .
Levels of Sharing for Platforms with Co-operative Infrastructures
Simulation Results for Enforced Fixed Contribution . . . . . . . .
Simulation Results for Voluntary Fixed Contribution . . . . . . . .
Simulation Results for Voluntary Variable Contribution . . . . . .
Voluntary Variable Contribution with mostly Selfish User Types .
Utility per User for Different Scenarios . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.

98
99
100
106
117
119
120
121
123

5.1.
5.2.
5.3.
5.4.
5.5.

Welfare Performance for Complete and Uncorrelated Preferences . .
Welfare Performance for Complete and Correlated Preferences . . . .
Fairness Performance for Complete and Uncorrelated Preferences . .
Fairness Performance for Complete and Correlated Preferences . . .
Algorithm Performance in Number of Matched Pairs, Uncorrelated
Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Algorithm Performance Relative to Optimum, Uncorrelated Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Performance of Matching Algorithms, Random Preferences . . . . . .
Comparison of Algorithms for Different Preference Lengths . . . . .
Algorithm Performance, Asymmetric Problem Instances . . . . . . .
Algorithm Performance, Correlated Preferences . . . . . . . . . . . .

164
166
167
168

5.6.
5.7.
5.8.
5.9.
5.10.

.
.
.
.
.
.
.
.
.

170
171
174
175
176
177

ix

x

List of Figures
6.1. Effects of Truncation Strategies on the Manipulating Users, GATA . . 192
6.2. Effects of Truncation Strategies on the Nonmanipulating Users, GATA 193
6.3. Average Gain and Number Benefiting Users for GA-Learning, 20x20
users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
6.4. Average Gain and Number Benefiting Users for PA-Learning, 20x20
users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
6.5. Number of Matched Pairs for Intermediate Supply and Demand . . . 204
6.6. Instability Effects for Intermediate Supply and Demand . . . . . . . . 205
6.7. Comparison of Matching Heuristics for Intermediate Supply and
Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
A.1.
A.2.
A.3.
A.4.
A.5.

Incentive Survey Logic . . . . . . . . . . . . . . . . . .
Platforms Used for Previous Online Resource Sharing
Resource Types Previously Shared Online . . . . . . .
User Groups with which Storage would be Shared . .
Required Relationships for Sharing Storage . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

237
238
238
238
239

List of Tables
2.1. Comparison of Simulation Tools . . . . . . . . . . . . . . . . . . . . .
2.2. Overview of Incentive Classification and Types . . . . . . . . . . . . .

36
41

3.1. Comparison of Survey TIPI Scores with original TIPI Scores and
TIPI-G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2. Relevance of Incentives for Private and Professional Networks . . . .

80
82

4.1.
4.2.
4.3.
4.4.

Model Parameters . . . . . . . . . . . . . . . . . . . . . . .
List of Utility Function Types . . . . . . . . . . . . . . . .
Incentive Scheme with Different User Type Distributions
List of User Cluster . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

. 87
. 90
. 102
. 110

5.1.
5.2.
5.3.
5.4.
5.5.

Computational Complexity of Two-Sided Matching Problems . . . .
Overview of Related Work . . . . . . . . . . . . . . . . . . . . . . . . .
Simulation Input Parameters . . . . . . . . . . . . . . . . . . . . . . .
Comparison of Algorithm Runtime . . . . . . . . . . . . . . . . . . . .
Algorithm Performance Relative to Optimal Solution, Incomplete
Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6. Algorithm Performance for Welfare and Fairness, Incomplete Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7. Algorithm Performance, Incomplete Correlated Preferences . . . . .
6.1. Absolute Preference Gain for Truncation Strategies, Different Truncation Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2. Absolute Preference Gain for Truncation Strategies, Different Number of Manipulating Users . . . . . . . . . . . . . . . . . . . . . . . . .
6.3. Effects of Manipulation on Stability . . . . . . . . . . . . . . . . . . . .
6.4. Manipulation Effectiveness . . . . . . . . . . . . . . . . . . . . . . . .

139
144
158
162
171
172
177
194
194
195
196

A.1. Spearman-Rho Correlation Table for Incentives for Previous Sharing,
Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
A.2. Spearman-Rho Correlation Table for Incentives for Previous Sharing,
Part 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
A.3. Spearman-Rho Correlation Table for Participation in Private Networks241
A.4. Spearman-Rho Correlation Table for Participation in Professional
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

xi

xii

List of Tables
B.1.
B.2.
B.3.
B.4.
C.1.
C.2.
C.3.
C.4.
C.5.
C.6.
C.7.
C.8.
C.9.

Simulation Results for Enforced Fixed Contribution . . . . . . . . . .
Simulation Results for Variable Fixed Contribution . . . . . . . . . . .
Simulation Results for Voluntary Variable Contribution . . . . . . . .
Simulation Results for Voluntary Variable Contribution, Selfish User
Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

243
244
245

Complete Preferences, Welfare Optimization . . . . . . . . . . . . .
Complete Preferences, Fairness Optimization . . . . . . . . . . . . .
Complete and Correlated Preferences, Welfare Optimization . . . .
Complete and Correlated Preferences, Fairness Optimization . . . .
Incomplete Preferences, 10x10 - 100x100 Users . . . . . . . . . . . .
Incomplete Preferences, 200x200 - 500x500 Users . . . . . . . . . . .
Incomplete Preferences, Asymmetric Sides, 10x50 - 20x100 Users . .
Incomplete Preferences, Asymmetric Sides, 50x10 and 50x100 Users
Incomplete and Correlated Preferences . . . . . . . . . . . . . . . . .

248
249
250
251
252
253
254
255
256

.
.
.
.
.
.
.
.
.

246

D.1. Absolute Preference Gain for Truncation Strategies, 1-10 Manipulating Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
D.2. Absolute Preference Gain for Truncation Strategies, 10-20 Manipulating Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259

List of Abbreviations
EFC
GA
GATA
IP
LP
P2P
QoS
RSMA
SC
SES
SLA
SM
SMT
SMTI
TA
TIPI
VC
VFC
VM
VVC

Enforced Fixed Contribution
Genetic Algorithm
Genetic Algorithm with subsequent Threshold Accepting Algorithm
Integer Program
Linear Program
Peer-to-Peer
Quality of Service
Requester-Optimal-Stable-Matching
Social Cloud
Social Exchange Simulator
Service Level Agreement
Stable Matching
Stable Matching with Ties
Stable Matching with Ties and Incompleteness
Threshold Accepting Algorithm
Ten Item Personality Inventory
Volunteer Computing
Voluntary Fixed Contribution
Virtual Machine
Voluntary Variable Contribution

xiii

List of Symbols
αi
β
χr,j,i (t)
δ
δi,r
ηr
i, j, k
κ1 , κ2
κT
λ
ωi,r
ωmin
o∈O
p
Πi,o
Πi,s
Pi
Pi,j
i

Availability of user i
Convexity parameter for utility function
Feedback of user j about user i for resource type r
Percentage of transactions requiring resources at the same time
Percentage of resources user i shares
Relative scarcity of resource type r
User indices
Weight of score si (t) components
Threshold value for scarcity
Degree of altruism
Initial resource endowment of user i and resource type r
Minimum number of provided resources
Outcome based on the set of possible outcomes O
Relative price for sharing resources
Provisioning of resources from user i
Consumption of resources from user i
Preference Profile of user i towards other users
Preference rank that user i has towards user j
Priority structure of user i, where i denotes strict priority and ∼ denotes indifference
r
Resource type
System requirements function
R (n)
si ( t )
Score of user i at time t
squantity,i (t) Score (Quantity) of user i at time t
sscarcity,i (t) Score (Scarcity) of user i at time t
Percentage of resources user i reserves for own purposes
σi,r
Minimum level of resources reserved for own usage
σmin
Strategy set for user i
Si
τi,r
Utility of user i for sharing resource type r
ϑ
Trust scores
Type of user i
θi

xv

xvi
Ui
ui (θi,o )

List of Abbreviations
Utility of user i
Utility of user i given its type θi and outcome o

Part I.
Introduction and Foundations

Chapter 1.
Introduction
“The Internet and especially Web 2.0 has brought about many new ways of sharing as well
as facilitating older forms of sharing on a larger scale.”
(Belk 2013)

R

ESOURCE allocation has always been one of the core foci of economic research. Whenever one group of people has access to, or can provide certain resources, and another group is interested in consuming these resources, (economic) mechanisms can facilitate such an exchange. Over the years, several types
of allocation mechanisms have emerged as principle ways to trade, share and exchange resources, such as fixed-price markets, auctions, and negotiations. In general, they can be categorized into mechanisms where the exchange involves monetary transactions, and non-monetary mechanisms.
The advent of computers had drastic effects on resource exchange mechanisms.
On one hand, the emergence of computer-based and computer-mediated exchange
platforms fundamentally changed the ways that resources are traded or shared,
affecting the speed, complexity and transparency of resource allocation in many
areas. This ranges from a rapid increase in speed through algorithmic traders on
stock markets to internet-based sharing platforms where cars1 , skills2 , travel accommodation3 , and other resources are shared between members of the platform.
On the other hand, besides facilitating the exchange, computing resources (storage,
computational power, etc.) represent a commodity that is becoming increasingly
1 http://www.zipcar.com/;

https://www.stadtmobil.de/ – last accessed May 2014
– last accessed May 2014
3 http://www.couchsurfing.com/ – last accessed May 2014
2 http://www.skillshare.com/

3

4

Introduction

important. As the generated data from research and businesses increases and computational models become more complex to solve, the demand for computational
resources changes accordingly. The evolution of distributed computing paradigms
such as Cluster Computing within an organization and Grid Computing across
organizations (Foster and Kesselman, 2003) reflect the necessity to provide access
to a larger set of computational resources. Over the last years, Cloud Computing emerged as the dominant paradigm to provide various types of computational
resources (such as computational power or storage) on-demand (Armbrust et al.,
2010). A multitude of different exchange mechanisms have been suggested and are
used for these paradigms. Especially when the resource exchange involves different organizations, monetary mechanisms are commonly used to gain access to the
resources (see Buyya and Murshed (2002) for Grid Computing and the large Cloud
Computing vendors such as Amazon Web Services4 and Rackspace5 ).
A prevalent economic argument for the Cloud Computing principle of on-demand
access to third-party computational resources is the “Capex to Opex” principle
(converting capital expenses to operating expenses; Armbrust et al. 2010, p.53).
Yet, it also exemplifies the change from a property-dominated paradigm in which
resource ownership was considered important, to an access- and service-based
paradigm (Rifkin, 2000; Vargo and Lusch, 2004). This paradigm change is in line
with the observation that the importance of the resource sharing concept, although
being far from new, has steadily increased over the past years and receives considerable renewed interest in both research and media (see e.g. Levine 2009; Belk 2010).
With the introduction of Web 2.0 technologies that facilitate the online communication and interaction between people, the last years have seen a rise in the number of
platforms that facilitate the sharing of certain resources. This ranges from the previously mentioned sharing of physical resources (cars, travel accommodation, etc.),
to digital or virtual goods as well as computation resources (files6 , storage7 , computational power8 , knowledge artifacts9 , etc.). Two important concepts for online
resource sharing in this context are Peer-to-Peer (P2P) (sharing) networks in which
resources are exchanged directly between users of the network in a decentralized

4 http://aws.amazon.com/

– last accessed May 2014
– last accessed May 2014
6 https://www.dropbox.com/ – last accessed May 2014
7 http://friendstore.news.cs.nyu.edu/ – last accessed May 2014
8 http://boinc.berkeley.edu – last accessed May 2014
9 http://www.myexperiment.org, http://nanohub.org – last accessed May 2014
5 http://www.rackspace.com/

Introduction

5

manner, and Volunteer Computing where computational resources are donated to
(scientific) projects.
As social connections and the trust inherent to them is a facilitator of economic exchange (Granovetter, 2005), a potential reason for this renewed interest in sharing
is the increasing prevalence and importance of (online) social networks in people’s everyday lives. This not only manifests itself in the rise of social network
platforms such as Facebook with over a billion users, but also in other (sociallydriven) Web 2.0 driven platforms (such as YouTube, Twitter, etc.) which thrive on
user interaction and contributions. Considering resource exchanges, the increasing
importance of social connections (in contrast to the mostly anonymous Web and
platforms that emerged in the early years of the Internet) also does not go unnoticed. The mentioned platforms for sharing physical or virtual goods are example
of platforms which leverage the social connections of users to facilitate resource exchanges. In contrast to monetary-based markets in which resource exchange necessarily involves certain types of payments, these new types of exchange platforms
often implement non-monetary based allocation mechanisms as a means to allocate resources. In particular, the social context might induce incentives such as
reciprocal or altruistic participation (Fehr and Schmidt, 2006).
Combining these aspects, and considering the increased interest in security aspects,
many exchange platforms begin to focus on social, non-anonymous resource sharing. Examples are P2P cloud computing (Babaoglu et al., 2012), the sharing of
computational storage among friends (Tran et al., 2008), and even the sharing of
insurance policies.10 Similarly, Swamynathan et al. (2008) proposes a “social marketplace” which combines traditional online marketplaces with social network features to enable purchases between friends and friends of friends.
This amalgamation of online resource exchange and social networks is also the
background of this work. The Social Cloud paradigm (Chard et al., 2012) serves as
the main use case of this thesis, and is explained in detail in Section 2.1. In particular, a Social Cloud (SC) is defined as the sharing of (computational) resources and
services between members of an underlying social network. The design of such
a platform is far from trivial, and several challenges need to be addressed for the
creation of a successful Social Cloud platform. These challenges are outlined in the
next section.

10 https://www.friendsurance.de/

– last accessed May 2014

6

Introduction

1.1. Design Challenges
A Social Cloud is a form of electronic marketplace where users share and exchange
resources. As such, there are several design decisions that have to be addressed:
the interaction of users with the platform, the mechanism to exchange resources,
the underlying technical infrastructure, etc. The methodology of Market Engineering (Weinhardt et al., 2003; Neumann, 2004) provides the means to structure this
design process. In particular, it focuses on the design of institutions to implement
and facilitate the resource exchange. Starting with the definition of the transaction
object of the market (i.e., the types of resources that are exchanged), Market Engineering considers several viewpoints of institutional design: the definition of a
market microstructure that specifies the economic mechanisms used to determine resource allocations and transactions; the specification of an IT infrastructure to host
the electronic marketplace and its functionality; and a business structure that addresses aspects such as business models and data ownership. The goal of Market
Engineering is to define these institutions such that they incentivize a certain user
behavior that results in a desired market outcome.
Using the Market Engineering approach, several challenges can be identified in the
design of a Social Cloud. This section provides an overview of the most relevant
design challenges, gives an extended description of the respective challenges and
specifies which of the challenges are addressed by this thesis.

Trust
In the context of resource sharing, or online cooperation in general, the existence
of trust is a fundamental factor that facilitates the exchange (Granovetter, 2005).
Arrow argues that “[v]irtually every commercial transaction has within itself an
element of trust, certainly any transaction conducted over a period of time. It can
be plausibly argued that much of the economic backwardness in the world can be
explained by the lack of mutual confidence” (Arrow, 1972, p.357).
Trust and its relation to exchange platforms has been studied in various disciplines.
Computer science mainly focuses on the application of trust concepts in platforms
for resource exchange. Examples are the calculation of trust scores between users of
a social network (Golbeck, 2005), and various reputation systems that aim at establishing the trustworthiness of users involved in resource exchanges (e.g., PeerTrust
(Xiong and Liu, 2004), PowerTrust (Zhou and Hwang, 2007), EigenTrust (Kamvar

1.1. DESIGN CHALLENGES

7

et al., 2003)). In economics, the concept of social or other-regarding preferences tries
to incorporate the effects of trust on user behavior in resource exchanges, mainly
through interpreting reciprocal actions as signs of trust between users (Berg et al.,
1995; Fehr and Schmidt, 2006). In sociology, trust is mainly perceived as calculated
risk taking in decision situations within social interactions (e.g., Coleman (1990)),
and several layers of trust are distinguished (Endress, 2012). From an interdisciplinary perspective, the roles of institution-based trust in the design of (online)
marketplaces are of importance (Pavlou and Gefen, 2004).
Whereas trust is modeled as a single score in the previously mentioned reputation systems, Endress (2012) argues that trust has multiple layers. There can be a
baseline trust between two users depending on their relationship, yet this can be
superimposed by context-specific trust, e.g., the perceived capability of providing
a certain outcome or action. It might also depend on the prior expectations of the
interacting users. Hence, trust is a multi-faceted concept that needs careful elaboration in the context of social resource sharing. Although first steps into this direction
have been taken by Caton et al. (2012), this challenge is not focus of this thesis.

Volatile Resources and User Availability
Social resource sharing differs from both ownership-based resource usage as well
as traditional payment-based resource usage. In contrast to resource ownership,
users who request resources have to take into account that resource availability
depends on the user providing the resource, e.g., if the resource is currently already
used, how long it can be used, etc. Also, especially for computational resources the
requested resources might not be as available and as reliable as resources provided
by professional companies, and often involve a best-effort type of usage policy.
For example, storage or computational power provided by users on a voluntary
basis depends on how much storage is available on their system, how long their
machines are online, etc.
User availability and the corresponding volatile resources have certain consequences for participating users. For example, if a user requests computational
resources to store data or use computational power, the unavailability of the providing user can cause serious consequences. This challenge is also relevant from a
technical point of view. If, for example, users exhibit a pattern of availability and
unavailability, and the platform wants to guarantee that data stored on other users’

8

Introduction

machines is available at any time (or with a certain probability), then techniques
such as data replication have to be considered.
This challenge is partially addressed in the thesis. In particular, the thesis studies
the effects of user availability and unavailability on the platform reliability in a
case where the resources required to run the platform are provided by the users
themselves.

Technical Implementation
Every online platform depends on a sound technical implementation to ensure reliable performance, achieve a high availability, and provide the necessary functionality for user management and user interaction. This basic functionality includes
databases to store relevant data about users, transactions, and other information,
as well as the definition and implementation of the user interface through which
users interact with the platform. From the perspective of a sharing community,
the platform also has to facilitate the actual exchange of (computational) resources.
This is a non-trivial technical challenge, as it requires the consideration of different
device technologies (e.g., mobile vs. static access), operating systems, and changing network addresses.
For a Social Cloud, Chard et al. (2012) and Caton et al. (2014) provide examples
how such a platform can be implemented, and Seuken et al. (2010) specifically considers user interface design for a computational storage sharing market. In this
thesis, the implementation challenge is partially addressed by the description of
an implemented Social Compute Cloud for sharing Virtual Machines. Usability aspects, on the other hand, are not the focus of this thesis and can be considered an
interesting opportunity for future work.

User Participation
Once the technical infrastructure for the platform is implemented, the next challenge is to attract users to participate on the platform. For online communities
in general, and sharing communities in particular, a commonly agreed upon challenge is to gain a critical mass of participating users (Markus, 1987; Van Slyke et al.,
2007; Westland, 2010) at which point network effects can be harnessed. With an
increasing number of users, the available resources and thus potential utility from

1.1. DESIGN CHALLENGES

9

using the platform generally increases as well. The design of appropriate incentives is specifically targeted to increase user participation. The potential motivations of users to join and participate on an online (sharing) platform can be diverse
and heterogeneous, and not all users can be expected to exhibit the same behavior
when presented with particular incentives. Hence, an incentive scheme targeted
at increasing the number of actively participating users has to be aware of the potentially heterogeneous user pool, and provide appropriate incentives (Vassileva,
2012).
Participation incentives encompass the first main part of this thesis. In the context
of Social Clouds, relevant motivations and incentives will be identified and discussed. Furthermore, two case studies are presented that highlight the usefulness
of a simulation-based approach in the design of incentive schemes.

Resource Allocation
Platforms for resource sharing or exchange are characterized by the fact that some
users have resources to offer that can be used or consumed by other users. In
such a context, the decision about the type of (economic) mechanism that facilitates
this exchange is fundamental. Platforms, in general, have a variety of options for
such mechanisms. One possibility is to allow users to contact, negotiate and share
resources in a decentralized manner, leaving the decision process which users to
contact solely to the users themselves. Another possibility is to provide a certain
form of centralized mechanism that determines the transactions and potential remunerations between users.
There are two main challenges in the design of a suitable resource allocation mechanism for Social Clouds, or online platforms in general. First, it has to be decided
which type of allocation mechanism is to be used, i.e., whether a decentralized
mechanism is used or a centralized mechanism with or without monetary transactions. Second, the chosen type of mechanism has to be adapted for the specific
scenario in order to guarantee certain desirable criteria, such as efficiency or fairness. Ideally, the allocation mechanism is congruent with the platform philosophy
and is chosen according to its ability to reflect what is important for the participating user types. For example, in social settings non-monetary mechanisms might be
preferred as monetary remunerations can affect the non-monetary motivations of
participating users (Frey and Jegen, 2001; Bénabou and Tirole, 2003).

10

Introduction

The design of a suitable resource allocation mechanism is the second main focus
of the thesis. For the first mentioned challenge, the findings of the relevant participation incentives are used to identify the suitable type of allocation mechanism.
Building on this, for the second challenge of designing the mechanism according
to the requirements of the platform, algorithms to find good allocations according
to several performance criteria are introduced and evaluated.

User Behavior
Online communities generally have an etiquette that users should not engage in
harmful or malicious activities on the platform. Especially when resources are
shared or exchanged, the tampering or misuse of the resources can have serious
negative effects on the resource providing users, the resources, as well as the platform itself (e.g., with respect to platform reputation). Besides misusing resources,
the strategic provisioning of false information, for example in resource exchanges,
might be beneficial for certain users if the respective exchange mechanism is not
incentive compatible. As such misrepresentation of information can be harmful
to other users and for the system as a whole, it is generally deemed desirable to
implement mechanisms that do not allow users to gain through information misrepresentation.
Although users might have to sign community guidelines when joining the platform, this does not ensure that the rules are actually followed. Different approaches
can be pursued to address this challenge. Through the design and provisioning of
certain incentives, malicious or harmful behavior can be made less appealing for
users. The design of formal agreements (similar to Service Level Agreements) is
used by certain platforms to lay out the terms of the resource exchange, and to
specify actions if the agreement is not adhered to. Furthermore, from a strategic
point of view, if the expected gain from malicious behavior or information misrepresentation is negative, then (rational) users have an (economic) incentive to
abstain from such behavior. Considering the technical implementation, additional
security measures can be used to minimize the risk of exploitation.
Several aspects of strategic or malicious behavior are discussed throughout the thesis. Specifically, in the context of resource allocations the thesis investigates if users
can gain from misrepresenting their private information.

1.2. RESEARCH OUTLINE

11

1.2. Research Outline
For the design of a sharing platform such as a Social Cloud, the previously mentioned challenges have to be addressed. This thesis contributes to this field by
focusing on two of these challenges: 1) User participation incentives, and 2) the
design of non-monetary resource allocation mechanisms.
Considering user participation incentives and contribution schemes, the thesis
aims to answer following research question:
R ESEARCH Q UESTION 1 ≺ PARTICIPATION AND C ONTRIBUTION I NCENTIVES 
What are relevant user participation incentives for Social Clouds, and how can they be
leveraged in the design of tailored participation and contribution schemes?
The first step in the design of appropriate and effective incentive schemes is the
identification of different participation stages that define how and to what degree
(potential) users interact with the platform. The existence of several distinct stages
with specific requirements for user integration and participation has already been
identified for general online communities (Jones and Rafaeli, 1999; Iriberri and
Leroy, 2009). This is an important consideration as the motivations and incentives might be different between stages. For example, the motivation to join a platform can be different from the motivation to actively contribute to the platform.
However, Social Clouds have unique, potentially different requirements for user
participation as other online communities. This stems from the fact that (bilateral)
resource exchange between users on the basis of existing social connections might
have an impact on the user behavior and the existing relationships. Therefore, research question 1.1 is posed to identify the relevant participation stages in which
the user interacts with the platform.
R ESEARCH Q UESTION 1.1 ≺ I NCENTIVE E NGINEERING  What are the stages of
participation and the corresponding relevant incentives that users exhibit in Social Clouds?
The results of research question 1.1 are necessary to design incentive schemes that
take into account the specific requirements of the different participation stages.
Chapter 3 addresses this research question by utilizing two approaches. On one
hand, through a comparison of related platforms and exchange systems as well as

12

Introduction

participation studies in online communities, a conceptual model of the participation stages is constructed. Additionally, the specific challenges within the separate
stages are discussed in the context of a Social Cloud. On the other hand, building
on the previously derived model, a small-scale web-based survey aims to identify the relevance of certain incentives within the mentioned participation stages.
In particular, its goal is to find relationships between user characteristics and the
relative importance of certain incentives.
By design, the introduction of an incentive scheme can affect how users interact
with other users and the system, for example by providing incentives to increase
one’s sharing activity. Before being applied in practice, the effects of such a (potentially novel) incentive scheme on the overall platform have to be predicted to ensure that the scheme achieves its goals. Often, analytic modeling is used to predict
said effects, or the scheme is introduced for a subgroup of users and subsequently
analyzed. However, both approaches are not always feasible. For example, a realistic analytical model might be too complex for formal evaluation, or the platform
for which the incentive scheme is designed might not exist yet. For these reasons,
this thesis proposes and applies a simulation-based approach as complementary
methodology to study potential effects of incentive schemes on the system. In particular, such an approach can yield predictions which are unattainable by the other
methods, such as dynamic effects on different user groups. To demonstrate how
such a simulation-based approach can augment the incentive scheme design, research question 1.2 studies the effects that can be predicted by such an approach.
R ESEARCH Q UESTION 1.2 ≺ I NCENTIVE S CHEME D ESIGN  How can a simulationbased approach be leveraged in the design of incentive schemes for participation?
As an example for the design of an incentive scheme, the introduction of a participation constraint in a Social Cloud is considered in a case study in Chapter 4.
The aim of such a participation constraint is to provide incentives for users to contribute resources to the system. It is well-known that users might have different
preferences to contribute resources and can be distinguished into different types
(Andreoni and Miller, 2002). Although the potential effects of the constraint on the
overall system have been studied before (see, e.g., Ranganathan et al. 2004), the
implications on different user types, and the consideration of system performance
depending on the distribution of such user types, is a novel scenario that requires
investigation. Furthermore, from a purely analytical approach it is not clear how

1.2. RESEARCH OUTLINE

13

the participation constraint will affect the resource contribution of the users dynamically. Hence, research question 1.2 investigates the effects of such an incentive scheme on different user groups and determines if the participation constraint
achieves its purpose.
User participation incentives, as studied in research questions 1.1 and 1.2, are necessary to create a sustainable platform with continuing and active user participation. From a technical point of view, such an online platform also requires certain technical infrastructure resources to host the platform itself, e.g., for services
such as user registration and management. Besides using dedicated third-party
resources for the infrastructure, it is also possible to let the users provide the necessary resources themselves. The concept of a co-operative infrastructure, in this case, is
defined as a platform where the resources that host the platform are provided and
owned by the users. As the feasibility of such an approach depends on the contribution of the users, an incentive scheme needs to be designed such that enough
resources are provided to ensure a certain platform availability. Research question 1.3 considers this scenario and focuses on the feasibility of different incentive
schemes.
R ESEARCH Q UESTION 1.3 ≺ C O - OPERATIVE I NFRASTRUCTURES  What are the
effects of different contribution schemes on co-operatively provided infrastructure resources
for Social Clouds?
To answer this research question, the previously described simulation-based approach is used in a second case study in Chapter 4 to examine several different
contribution schemes for users. For example, a certain contribution to the infrastructure might be enforced for all users who want to participate, or the decision
to participate can be completely left to the users themselves. Research question
1.3 studies how different contribution schemes and assumptions about user type
distributions affect the feasibility of such a co-operative approach.
The first main research question considers aspects of user participation and resource contribution, which was identified as one of the key challenges in the design
of a Social Cloud platform. Given that users participate on the platform and provide resources, the second key challenge is to find appropriate (matching) mechanisms that allocate resources in line with the goals of the platform. As discussed,
in many economic settings the allocation of resources involves monetary transac-

14

Introduction

tions based on (private) valuations that the market participants have for the resources. While such monetary-based mechanisms might be useful for many exchange settings, there are also scenarios for which non-monetary mechanisms are
considered more useful. This is especially the case in settings where monetarybased exchanges would be considered unsocial, unethical or illegal. Examples of
such settings are the matching of children to schools (School Choice Problem, see
e.g. Abdulkadiroğlu and Sönmez, 2003) and the matching of college students to
college spaces (College Admission Problem, see e.g. Roth and Sotomayor, 1992).
Considering resource sharing platforms, especially ones with a social setting such
as Social Clouds, the use of non-monetary mechanisms can be observed frequently.
Examples are the use of credits and trophies on nanoHUB.org11 and myExperiment.org12 (called “nanos” and “reputation points”, respectively) for sharing research artifacts as well as learning and teaching materials; P2P platforms where
users can exchange electronic goods; or storage-sharing platforms where users can
supply and use storage provided by other users without monetary compensation
(see e.g. Seuken et al., 2010). In these platforms, participants (which can still be distinguished into resource providers and resource consumers) share and consume
available resources without the immediate goal of monetary gain. A characteristic
of such platforms is that in contrast to mostly anonymous platforms, users of such
“social” platforms often value non-monetary incentives (such as reciprocity, altruism, etc.) higher than purely monetary remuneration when it comes to the sharing
and exchange of resources (Bénabou and Tirole, 2006; Fehr and Schmidt, 2006).
In such a non-monetary setting, the theory of two-sided matching is well established
as a means to allocate resources. The key aspect of two-sided matching is that
users specify a preference ranking with whom they want to share and exchange
their resources. Based on these preferences, a matching mechanism tries to find an
allocation with certain properties (Roth, 2008). Depending on the structure of the
preferences and the desired properties, several algorithms have been developed
to compute such an allocation. However, there are some drawbacks with existing
solutions: 1) there are scenarios for which the calculation of an optimal solution
is NP-hard and no suitable approximation algorithms with a guaranteed quality
bound exist (Halldórsson et al., 2003). 2) Existing algorithms are all developed for
a given scenario and a certain combination of goals, thereby lacking the flexibility of being applicable in different scenarios. 3) Existing algorithms concentrate
11 http://nanohub.org

– last accessed May 2014
– last accessed May 2014

12 http://www.myexperiment.org

1.2. RESEARCH OUTLINE

15

on achieving a stable solution, which infers that no user can be better of by deviating from the given solution; due to an impossibility results (Roth, 1982), the
corresponding mechanisms are not incentive compatible, and the effects of strategic
manipulation have to be considered. 4) Two-sided matching algorithms are often
considered in a static context, where the allocation is calculated in a batch-like procedure, which does not reflect potential dynamics of a real platform such as a Social
Cloud.
This is the focus of the second part of the thesis, which aims to address these
drawbacks and considers algorithms for preference-based resource allocation13 as
a means to combine non-monetary mechanisms with the advantages of centralized,
market-based allocation:
R ESEARCH Q UESTION 2 ≺ R ESOURCE A LLOCATION  Which types of algorithms
provide a good combination of performance, flexibility, and strategic properties for
preference-based resource allocation?
As the allocation mechanism should have a certain flexibility to adjust to different goals and scenarios, heuristics have been suggested to calculate allocations in
preference-based matching (Vien and Chung, 2006; Kimbrough and Kuo, 2010).
This thesis follows a similar direction and extends this work by developing heuristics that have the ability to handle a variety of preference structures and goal combinations. As there might be a trade-off between the flexibility and the achieved solution quality of such heuristics compared to existing algorithms, the performance
of the heuristics is the focus of research question 2.1:
R ESEARCH Q UESTION 2.1 ≺ P ERFORMANCE OF H EURISTICS  What is the per for mance of heuristics for preference-based matching compared to existing matching
mechanisms?
Chapter 5 compares the performance of different algorithms and heuristics in several standard scenarios. As heuristics allow for more flexibility over existing algorithms with respect to goals and preference structures, an equal or improved
performance with respect to certain goal metrics would show the general applica-

13 In

the remainder of the thesis, preference-based matching and two-sided matching will be used
interchangeably.

16

Introduction

bility of said heuristics and provide a valuable contribution to the field of two-sided
matching.
Besides pure performance and solution quality characteristics, there are additional
considerations that are of practical interest in preference-based matching. One
such aspect is the consideration of strategic behavior in a market. From the point
of view of participating users, the question of how to interact with the matching
mechanism arises. As the mechanisms calculate solutions based on the preference
rankings that are provided by the users, they might have an incentive to manipulate their submitted preferences in the hope to benefit from manipulation. In fact,
the impossibility result of Roth (1982) shows that for the standard set of goals in
preference-based matching (if a stable solution has to be guaranteed), there can
be no incentive compatible mechanism, i.e., for which it is the best strategy for
all users to submit their true, non-manipulated preferences. This also applies for
heuristics if a stable solution is of interest.
In general, the manipulation of the submitted preferences can have several effects.
Besides the direct effects on the manipulating as well as non-manipulating users,
which can either gain or lose by such manipulation, it might also change the quality of the solution with respect to the true preferences of users. For example, an
optimal solution with respect to the submitted preferences might not be optimal
under the true, non-manipulated preferences, and the effects of manipulation on
solution quality is a largely unexplored field. Hence, the focus of research question
2.2 is the potential effects of such manipulation:
R ESEARCH Q UESTION 2.2 ≺ I NCENTIVE C OMPATIBILITY  What are the effects of
preference manipulation on the manipulating users, non-manipulating users, and the solution quality?
Chapter 6.1 studies the effects of manipulation for considered preference-based
matching algorithms. In addition, it also considers the robustness of certain algorithms against potential manipulation. This aspect is particularly interesting if
participating users want to learn a beneficial manipulation strategy. Both the likelihood of a successful manipulation, as well as the average gains from manipulation
are considered.
Besides flexibility, performance, and incentive compatibility, the fourth aspect that
needs to be addressed is the general setting of the matching calculation. Preference-

1.3. STRUCTURE OF THE THESIS

17

based matching mechanisms usually assume that the allocation is calculated in a
batch-like procedure, where participants submit their preferences to the algorithm
once and are matched according to the resulting solution. In realistic settings, however, the scenario can be more complex. Instead of being matched only once, or at
certain time intervals, requesting and providing users might join and leave the system dynamically. This creates situations where new supply and demand arrives
in between the allocation time-slots. Research question 2.3 targets this issue and
studies options to deal with such dynamic supply and demand.
R ESEARCH Q UESTION 2.3 ≺ D YNAMIC A LLOCATIONS  What are options to allocate dynamic supply and demand, taking into account potential existing matches?
Two straightforward strategies to deal with such dynamic supply and demand are
either recalculating the entire solution (taking into account the currently matched
users), or leaving the new supply and demand unallocated until the next allocation time-slot. The first option might not always be possible for time- or technical
constraints, e.g., if the calculation of the solution takes too long or breaking up
matched pairs is technically not feasible. The latter option, however, will potentially leave a considerable amount of available resource idle. Therefore, research
question 2.3 studies the effects of such dynamic supply and demand on preferencebased resource allocation. Besides the two mentioned strategies, two heuristics are
considered as an approach to match otherwise idle resources without breaking up
currently matched users.

1.3. Structure of the Thesis
The research outline as described in the previous section reflects the structure of the
thesis, which is comprised of four parts. Part I introduces the necessary concepts
as well as the main use case and methodology which is applied in the subsequent
parts. The design of participation incentive schemes is the focus of Part II, and Part
III considers mechanisms for resource allocation. Part IV concludes the thesis and
highlights future research directions.
A high-level illustration of this thesis’ structure is shown in Figure 1.1. Chapter 2
lays the foundation for the subsequent chapters by establishing the common terminology and main concepts used throughout the thesis. In particular, Section 2.1

18

Introduction

3DUW,

,1752'8&7,21 
)281'$7,216

3DUW,,

3$57,&,3$7,21
,1&(17,9(6

3DUW,,,

5(6285&(
$//2&$7,21

3DUW,9

),1$/(

Chapter 1
Introduction
Chapter 2
Resource Sharing in Social Contexts

Chapter 3
Incentives in Social Clouds
Chapter 4
Designing Incentive Schemes and Co-operative Infrastructures

Chapter 5
Resource Allocation in Social Clouds
Chapter 6
Incentive Compatibility and Dynamic Allocation

Chapter 7
Conclusion

Figure 1.1.: Structure of this Thesis

describes the concept of Social Clouds and provides details on an implemented
prototype thereof. Section 2.2 discusses a simulation-based approach and the corresponding simulation tool as main methodology in this thesis. The previously
mentioned challenges of user participation and resource allocation are further detailed in Section 2.3.
Building on these foundations, Chapter 3 tackles the challenge of providing incentives for user participation by identifying different stages of user interaction and
participation with the platform. Additionally, the results of a web-based survey
on the importance of different incentives within these stages and their dependency
on factors such as user type are presented. Chapter 4 consists of two case studies:
the first case study investigates the effects of a participation constraint on system
and user behavior and the dependency on user characteristics. The co-operative
provisioning of infrastructure resources is focus of the second case study, which
compares different contribution schemes and their effectiveness based on the distribution assumptions of user types.
In the third part of this thesis, Chapter 5 considers preference-based matching
mechanisms as a means to allocate resources in Social Clouds. Starting with an
outline of the main concepts and existing algorithms, it describes the proposed
heuristics and evaluates their performance for different preference structures and
goal combinations. Chapter 6 presents two additional topics of interest in the area

1.4. RESEARCH DEVELOPMENT

19

of preference-based matching. First, the effects of strategic manipulation of preferences are evaluated and discussed. Second, approaches to handle dynamic allocations are considered. Chapter 7 summarizes the key contributions of this thesis,
provides an outlook on future research and highlights complementary topics.

1.4. Research Development
Parts of this thesis have been presented and published at four peer-reviewed international conferences and workshops, as well as an international journal. This
section provides an overview of the published material and simultaneously outlines the development of the work and the corresponding refinement and extension
steps.
The identification and discussion of different stages of user participation during
their interaction with a Social Cloud (Section 3.2) was presented in a workshop at
the 11th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGrid 2011) (Haas et al., 2011).
The simulation tool described in Section 2.2 which is used to facilitate the
simulation-based evaluations in the thesis, along with the case study about the introduction of a participation constraint and the corresponding evaluation (Section
4.2) were presented at the 8th IEEE International Conference on eScience (eScience
2012) (Haas et al., 2012). The description and implementation of the simulation
tool was extended to capture additional scenarios, such as simulating preferencebased resource exchanges. Furthermore, the case study as presented in Section
4.2 additionally considers the comparison of different user type distributions. The
case study on co-operative infrastructures was presented at the 48th Hawaii International Conference on System Sciences (HICSS’48) (Haas et al., 2013). The corresponding Section 4.3 encompasses an additional contribution scheme that was
added as an extension for this thesis.
Considering the evaluation of preference-based matching heuristics and their comparison to existing algorithms, the results for the first scenario (Section 5.4.3) were
presented at the 10th International Conference on Economics of Grids, Clouds, Systems and Services (GECON 2013) (Haas et al., 2013). The evaluation was then extended to cover additional scenarios as well. The Social Compute Cloud prototype
as described in Section 2.1.4, as well as the consideration of dynamic allocations

20

Introduction

in the context of preference-based matching in Section 6.2, was published in IEEE
Transactions on Services Computing (Caton et al., 2014).

Chapter 2.
Resource Sharing in Social Contexts
“The recent changes in our economic landscape have notably exposed and intensified a
phenomenon: an explosion in sharing, bartering, lending, trading, renting, gifting, and
swapping.”
(Botsman and Rogers, 2010)

T

HIS chapter introduces the concepts and methodologies used throughout the
thesis. Section 2.1 introduces the concept of Social Clouds, a resource sharing paradigm in which resources are shared between members of existing social
networks. The Social Cloud concept will be used as unified use case throughout
the thesis. Additionally, a prototype of a Social Cloud for computational resources,
along with its implementation details, is presented. In conjunction with the prototype, a simulation tool was developed as complementary methodology to study
various aspects of a Social Cloud. Section 2.2 describes the simulation approach
and corresponding tool that is used as the main evaluation methodology in this
thesis. Applying literature reviews and a comparison to similar systems, Section
2.3 discusses two coordination challenges that have to be addressed in the design of
such platforms. In particular the consideration of user participation incentives, as
well as the type of (economic) resource allocation mechanisms, are identified as the
key challenges which this work addresses. Finally, Section 2.4 closes the chapter
with a summary of the described concepts.

21

22

Resource Sharing in Social Contexts

2.1. The Concept of Social Clouds
This section describes the concept of a Social Cloud (Chard et al., 2010, 2012) as
an example of a social resource sharing platform. It will be used as a unifying use
case that shows the practical application of the mentioned topics and will serve
as an example sharing platform throughout this work.1 Section 2.1.1 defines the
concept, describes the components and discusses use cases for the application of a
Social Cloud. Section 2.1.2 provides an overview of related concepts, and Section
2.1.3 discusses challenges in the design and construction of a Social Cloud. Section 2.1.4 focuses on a prototype system of a Social Compute Cloud and provides
implementation details.

2.1.1. Definitions
A Social Cloud is a dynamic environment through which (new) Cloud-like provisioning scenarios can be established based upon the implicit levels of trust that
transcend inter-personal relationships digitally encoded within a social network
(Chard et al., 2010, 2012). The concept of a Social Cloud is defined as:
Definition 1 ( Social Cloud, Chard et al., 2012 ). A Social Cloud is a resource and
service sharing framework utilizing relationships established between members of a social
network.
The vision of a Social Cloud is motivated by the need of individuals or groups
to access specific resources they are not in possession of, but that can be made
available by connected peers. In simple words, Social Clouds use social networks
as mechanisms for collaboration and resource sharing. Moreover, Social Clouds
rely on social incentives to motivate sharing and non-malicious behavior, as users
leverage their existing networks to share capabilities and resources (Haas et al.,
2013). A Social Cloud is a form of Community Cloud2 , as the resources are owned,
provided and consumed by members of a social community.
1 Therefore,

the terms “platform” and “Social Cloud” are used interchangeably.
defines a Community Cloud as: “[...] The cloud infrastructure is provisioned for exclusive use by a specific community of consumers from organizations that have shared concerns
(e.g., mission, security requirements, policy, and compliance considerations).” (Mell and Grance,
2011)

2 NIST

2.1. THE CONCEPT OF SOCIAL CLOUDS

23

There are several characteristics that distinguish a Social Cloud from other computing approaches. In contrast to Volunteer Computing (VC), which is defined as “a
form of distributed computing in which the general public volunteers processing
and storage resources to computing projects” (Anderson and Fedak, 2006, p.73),
users in a Social Cloud cannot only donate but also consume different resources
in exchange for their resource provision. Furthermore, contrary to traditional P2P
resource sharing which “allow a distributed community of users to share resources
in the form of information, digital content, storage space, or processing capacity”
(Krishnan et al., 2006, p.32), the resources are provided by users with direct or indirect relationships in a social network. Hence, resources are no longer offered by
anonymous providers but by socially connected users, where the existing relationships can be used to deduce some form of bi-lateral association or understanding
of trust. Furthermore, it is also different from the definition of Cloud Computing:
Definition 2 ( Cloud Computing, Mell and Grance 2011 ). Cloud computing is a
model for enabling ubiquitous, convenient, on-demand network access to a shared pool of
configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or
service provider interaction.
Whereas the NIST definition of Cloud Computing focuses on technical aspects such
as on-demand access, rapid provisioning and configurable resources, the Social
Cloud paradigm focuses on the utilization of the underlying social network for the
sharing of a pool of resources. Common aspects between the definitions are having
such a pool of different types of resources and services, and having (ideally) low
management effort and interaction.
Another key aspect in the definition of a Social Cloud is the existence of an underlying social network between the users. This is crucial as it has implications
about which mechanisms are used to facilitate resource sharing in such a system.
As Granovetter (2005) shows, social networks have an impact on market outcomes,
and trust emerges on the basis of a social network. Hence, the design of a Social
Cloud has to take the social network of users into account. This is elaborated in
more detail in Chapter 3.
Figure 2.1 shows an example SC where users provide (potentially heterogeneous)
resources, and a clearinghouse determines the allocation of supply and demand.

24

Resource Sharing in Social Contexts
Social Cloud

Clearinghouse
4

Demand

1

Supply

3

2

Figure 2.1.: Example of a Social Cloud

The connections between users show the underlying social network, and the numbered users indicate that these users are part of the Social Cloud. Note that not only
the resources, but also the types of users, in a SC can be heterogeneous. Different
types of users have different resource capabilities and requirements. The potential
heterogeneity of users and resources also affects the interaction of the users with the
SC clearinghouse as well as the appropriate choice of a market mechanism (which
is not restricted to preference-based matching for general Social Clouds).

Architecture
The general technical architecture of a Social Cloud is shown in Figure 2.2. The
main component is the Social Cloud platform, which consists of several modules.
The Clearinghouse module includes the allocation mechanisms and relevant corresponding social and economic protocols (e.g., social service level agreements that
provide a formal description of the exchanges). The Platform Manager provides
administration functionality and is responsible for the tasks such as user management. Several databases are required to store relevant data about user resources,
allocations, and user preferences. For the technical facilitation, a Resource Middleware component enables the execution of the Social Cloud platform on computational resources. The Social Cloud platform has two main interfaces. The
socio-technical adapter enables the extraction of social network data, and the user
interface defines how users communicate with the platform, e.g., through the specification of sharing preferences.

2.1. THE CONCEPT OF SOCIAL CLOUDS

25

Figure 2.2.: General Architecture of a Social Cloud

Application Scenarios for Social Clouds
The definition of a Social Cloud is flexible and allows for a range of potential application scenarios. Based on Chard et al. (2012), the following list describes the
potential scenarios.
Social Storage Cloud In this scenario, users of a Social Cloud share storage space,
for example to backup data, share pictures, and store large data sets. Given
certain security requirements, this scenario is perhaps the simplest to implement. A prototype of a Social Storage Cloud implemented as a Facebook
application has already been developed in (Chard et al., 2012).
Social Compute Cloud Instead of storage, (virtualized) computational resources
are shared in this scenario, which targets the potentially large un- or underused computing potential of personal computers. Such computational power
can be used by other members of the Social Cloud, or provided to scientific
communities for complex computations. This scenario is considered by John
et al. (2011), who study different incentive schemes to encourage contribution
to such public research projects. A prototype of a Social Compute Cloud is
also presented in detail in Section 2.1.4.
Social Collaboration Cloud The mentioned platforms myExperiment.org and
nanoHUB.org are examples for this type of scenario. Here, members of a
Social Cloud participate due to some common collaborative task or goal, for
example the sharing of knowledge artifacts within a scientific community. A
potential use case is the sharing of large datasets within a scientific community using a social content delivery network (Chard et al., 2012).

26

Resource Sharing in Social Contexts

Social Cloud for Public Research Closely related to Social Compute Clouds, this
scenario considers Social Clouds for providing computational resources to
Volunteer Computing-like projects. Specifically, not only could the provided
computational power due to unutilized resources be potentially increased
dramatically, research projects can also be propagated through the network
to friends and other members in- and outside the network, thereby increasing the potential user base (Thaufeeg et al., 2011).
Social Enterprise Cloud In contrast to the previous scenarios which focus on private users, enterprises can also leverage specialized Social Clouds through
their social networks. This would allow them to gain access to additional
computational resources, and at the same time increase the awareness of the
enterprise/company with participating users.

Trust
One of the fundamental assumptions of the Social Cloud paradigm is the existence
of a certain level of trust between the members of the social network. As mentioned
in Section 1.1, however, different scientific disciplines have their own version of
what constitutes the notion of trust, which makes an interdisciplinary focus necessary. For this reason, Caton et al. (2012) define trust in the context of a Social Cloud
from the perspectives of economics, computer science and sociology.
Definition 3 ( Trust in Social Cloud, Caton et al., 2012 ). Trust is a positive expectation or assumption on future outcomes that results from proven contextualized personal
interaction histories corresponding to conventional relationship types and can be leveraged
by formal and informal rules and conventions within a Social Cloud to facilitate as well as
influence the scope of collaborative exchange.
There are several important concepts addressed by this definition. The notion of
positive expectation on future outcomes is related to economic aspects such as expected reciprocity, where another user is trusted to return a favor in the future.
Proven contextualized personal interaction means that the specific trust depends on
the prior social context of previous interactions, e.g., indicated by the relationship
type between the users. Formal and informal rules and conventions is another relevant aspect which indicates that trust can be facilitated by explicit agreements (e.g.,
terms of collaboration), or informal norms that are enforced by the community.

2.1. THE CONCEPT OF SOCIAL CLOUDS

27

2.1.2. Related Concepts
There are several concepts similar to a Social Cloud as previously defined. Using
the same term, Mohaisen et al. (2011) propose a paradigm for trustworthy distributed computing on social networks. In the context of federated Cloud systems,
Elnaffar et al. (2013) describe a model of a social network where several cloud systems form different types of relationships with other clouds. Kourtellis et al. (2010)
present a P2P based service (“Prometheus”) that collects social information from
different sources and allows social-based mapping and to draw social inferences.
Kourtellis (2012) argue that the vast amount of information present in social network (platforms) can be leveraged to build socially-aware distributed systems with
various application scenarios, increasing aspects such as system response time.
From a technical point of view, several distributed computing approaches such as
ASPEN (Curry et al., 2008) and PolarGrid (Guo et al., 2009) leverage information of
social networks in their applications. OpenSocial (Häsel, 2011) provides interface
specifications that allow access to social network information in order to create
social web applications. P2P clouds (Ranjan et al., 2010) aim to create (private)
clouds on the basis of peer connections, yet do not have the same assumptions
about the underlying social network and the resulting trust between members as
in the concept of Social Clouds.
In particular (academic) domains, social networks are also used to build and coordinate special-interest communities. Prominent examples are the previously mentioned platforms myExperiment (De Roure et al., 2009) and nanoHUB.org (Klimeck
et al., 2008). myExperiment allows users to share (scientific) workflows, in order to determine commonalities and increase the dissemination of popular workflows (e.g., for scientific experiments) to the scientific community. nanoHUB.org
is a large, world-wide community of researchers in nanotechnology, and users can
share teaching and research material among each other.
Another related field is Volunteer Computing through systems like BOINC (Anderson, 2004). In projects such as SETI@home (Anderson et al., 2002), users donate computational resources to scientific projects which use the donated resources
to run extensive calculations. In contrast to Social Clouds, Volunteer Computing
is inherently unilateral as the concept does not involve the notion of bilateral exchanges. Furthermore, the participating users do not have to belong to the same

28

Resource Sharing in Social Contexts

social network and can remain completely anonymous, and social connections and
relationships are not leveraged.
A prominent example of sharing computational resources is the exchange of electronic storage between participating users. Examples of such a scenario are platforms like FriendStore3 (Tran et al., 2008), F2Box (Gracia-Tinedo et al., 2012), symform4 , P2P storage sharing (Seuken et al., 2010), or a Social Storage Cloud (Chard
et al., 2010) in which registered users can provide free storage space on their own
machines for data backup and other purposes to other users. In return, the contributing users are able to use the storage of other users for their own backup. Such
systems are alternatives to third-party storage and backup solutions. Security measures such as encryption algorithms or sandboxing technologies are important in
this scenario to avoid unauthorized access of one’s data on other users’ machines,
as well as protecting storage-providing users from malicious data.

2.1.3. Design Challenges
There are many challenges in the construction of a Social Cloud that need to be
carefully considered. The following discussion considers four key challenges: the
technical facilitation of the cloud platform, the inclusion as well as interpretation
of social (network) structures, the design and implementation of appropriate socioeconomic models for the facilitation of exchange, and the sustainability of the platform infrastructure.
Technical Facilitation To facilitate the exchange of resources, i.e., both provision
and consumption, the necessary technical infrastructure has to be provided.
This includes the handling of different types of users (e.g., mobile users with
non-static IP addresses) and resources (e.g., computational storage and Virtual Machines). In addition, even though the Social Cloud paradigm assumes a certain level of trust between users, certain security mechanisms
such as sandboxing are required. Such security mechanisms aim at protecting resources from potentially malicious users and also protect user applications from potentially malicious resources. These aspects can be partially
addressed by virtualization, which also enables the support for different operating systems.
3 http://friendstore.news.cs.nyu.edu/
4 http://www.symform.com/

– last accessed May 2014
– last accessed May 2014

2.1. THE CONCEPT OF SOCIAL CLOUDS

29

Leveraging Social Structures Users of a Social Cloud are connected through an
underlying social network. To utilize the underlying social network structure
for resource sharing, Social Clouds rely on users to allow access to their social
network and trust the platform with the corresponding data. The data that is
accessible on many social network platforms, which often utilize a binary representation of a “friend” relationship, might not be adequate to capture the
multitude and complexity of the specific relationships. Besides the relationship role, other factors such as (perceived) competency, the exchange context
(i.e., being a provider or consumer), and trust also influence the relationships
between users. Hence, a Social Cloud should provide the ability to augment
standard social graphs with additional (meta-)data, which can then be used
to model relationships more adequately and to potentially extract preferences
of users towards one another.
Socio-economic Model The applied resource sharing mechanisms in Social
Clouds have to consider the application scenario as well as the relationships
between users. Depending on the scenario, different mechanisms might be
suitable, and the Social Cloud should provide the necessary functionality to
enable the respective mechanism. In non-enterprise settings, where the focus
is on sharing rather than sale of resources, non-monetary mechanisms (in particular, preference-based matching) promise to align the social setting of the
platform with the advantages of market-based allocation of resources such
as allocation efficiency and stability.5 Congruent to the previous challenge,
a Social Cloud has to provide the means for users to express their sharing
preferences.
Platform Facilitation A Social Cloud needs certain computational resources to
provide a basic set of functionality (such as user management) and to facilitate the actual resource exchange. The use of third-party resources to host the
platform would require a revenue model (or similar) to cover the expenses,
and might contradict the concept of voluntary resource sharing. An option to
alleviate this challenge is to follow a co-operative model which implies that
users provide resources not only for sharing, but also to support the platform
itself in the form of infrastructure resources. Such a co-operative infrastructure model is introduced in Chapter 4.3.
5 This is the focus of the second part of this thesis.

of preference-based sharing.

Chapters 5 and 6 introduce and evaluate concepts

30

Resource Sharing in Social Contexts

2.1.4. A Prototype Social Compute Cloud
In order to demonstrate the usefulness and viability of a Social Cloud, this section
describes a prototype of a Social Compute Cloud in which participating users can
share virtualized computing resources with friends of their social network (Caton
et al., 2014). A Social Compute Cloud is designed to enable access to compute capabilities in the form of Virtual Machines provided by socially connected peers, and
users are able to execute programs on these contributed, virtualized resources.

Architecture and Implementation of a Social Compute Cloud
Building on the previously described general architecture of a Social Cloud (Figure
2.2), and addressing the challenges mentioned before, Figure 2.3 shows the architecture of the Social Compute Cloud and its core components. The Social Cloud
Platform provides the technical implementation details which are needed to facilitate the resource exchange. To enable the actual sharing of virtualized resources,
the Seattle framework is used as resource middleware component as it largely provides the needed functionality such as sandboxing (Cappos et al., 2009; Zhuang
et al., 2013). Seattle is an open source P2P platform designed to create a distributed
testbed to easily create distributed applications. It was chosen as the basis for the
implementation due to its lightweight virtualization middleware that is used to
enable application execution on contributed resources, and its extensible clearinghouse model.
Considering the socio-economic model that governs the resource exchange, a social clearinghouse encompasses the implementation of preference-based matching
algorithms, as well as a module to capture the user preferences. A new social clearinghouse that leverages these connections is implemented as the original Seattle
clearinghouse did not consider social connections and did not provide the functionality for preference-based allocation of resources. The implemented algorithms
are the matching algorithms described in Chapter 5. For the specification of sharing
preferences, the implementation provides an interface that allows users to specify
a preference rank for each user in their social network that is also part of the Social Compute Cloud. Given the outcome of the matching algorithms, users are
assigned to contributed and available resources. For the access to underlying social network, the platform uses a Facebook application to access a users’ social
network data (as well as providing a means for authentication). The platform also

2.2. SIMULATING SOCIAL CLOUDS

31

provides databases to store social network data as well as data about the provided
and requested resources.
Social CloudPlatform

SocialNetwork

Preferences Module
Matching /Allocation
Mechanism(s)
SocialDB

Sharing
Preferences

User

Resource
Request(s)

ResourceDB

Consume
Resources

ComputeResources

ProvideResources

Figure 2.3.: Architecture of a Social Compute Cloud and its Core Components

The prototype has been evaluated in a proof-of-concept study in (Caton et al., 2014).
As a next step, the prototype should be studied as a test system with real users. The
feedback of the deployed system would then serve to potentially improve the implementation and design of the Social Compute Cloud, and also to get data about
how users interact with the system and other users. Such feedback can provide additional insights in actual user behavior, for example considering their incentives
to participate on the platform and the process how they specify their preferences.
A complementary methodology to study Social Clouds is to use a simulation-based
approach. For this reason, the next section introduces a simulation tool that is used
throughout the thesis to model, simulate and analyze aspects of a Social Cloud.

2.2. Simulating Social Clouds
As Social Cloud scenarios begin to become more commonplace, it is important to
understand the intricacies of the corresponding exchange system. A primary example is the behavior of the system as a product of its users’ actions and interactions,
as these define how the system performs. This is important, as resource exchange
systems naturally include a high level of complexity stemming from the dependencies between the decisions of single users and the behavior of their peers and
extended peers, i.e. friends of friends. This challenge is aggravated by the different relationship types that can exist between users and the social context that is

32

Resource Sharing in Social Contexts

attributed to a given relationship type by a user. These contexts can also be different between users, but ultimately mean that a user’s behavior strongly depends on
the relationship types that exist within their resource sharing community and the
users they interact with. Therefore, in order to adequately design Social Clouds,
many experiments on user behavior and user management are required. Similarly,
in the wider context of computationally supported science (e-Science), a tool that
can help in the design and study of resource exchange communities will aid in
the understanding of their dynamics, but also in the design of new mechanisms to
improve different parts of their performance.
Given the social setting and myriad of potential user profiles for Social Cloud platforms, studying system behavior with a purely analytical approach may be inflexible or even unsolvable. Instead, a simulation environment can be a suitable
methodology to study the behavior of complex systems. Doing so will allow not
only the testing and analysis of new aspects of these systems. It also enables the
investigation of which aspects of underlying social circles and communities are important in their design, and the identification of methods to capitalize upon these
observations. A key opportunity that such a simulation tool opens up is the ability
to merge observations of exchange situations from existing systems or lab experiments of usage. A simulation tool can help not only in the design of such systems,
but also in identifying and pre-screening case studies and experiments to observe
real users as well as further analysis of completed user experiments to test and
form hypotheses about resource exchange in social settings. In other words, a simulation approach is a useful complementary methodology in the study of dynamic,
complex resource sharing systems.
This section provides details of the simulation tool (Social Exchange Simulator
(SES)) developed for this task. Application scenarios are described in Section 2.2.1,
and Section 2.2.2 provides a discussion of related work. Finally, Section 2.2.3 describes the architecture of the simulation tool. The main parts of this section have
been published in (Haas et al., 2012).

2.2.1. Purpose and Potential Applications
In the context of engineering, testing and introducing Social Cloud, the designer
has to anticipate the behavior of the platform under certain conditions. Several
standard methodologies are applicable for this task in general.

2.2. SIMULATING SOCIAL CLOUDS

33

1. Theoretical analysis of system properties.
2. Prototypical implementation and simulation of expected system behavior.
3. Empirical observation of user and system behavior.
The simulator targets the second methodology: based on certain assumptions
about user intelligence, behavior and certain system properties, it is able to study
dynamic system behavior and, for example, the predicted response of users to certain changes on the system rules. It is necessary to emphasize that such a simulation tool derives its value by being complementary to other research methodologies. It cannot substitute the insights a system designer can get from using the other
methodologies, but rather provide additional, complementary results that are helpful in the design of a resource sharing platform. For example, the simulation tool
can be used in a system that is complex enough that analytical predictions are not
feasible and at the same time is not implemented, leading to a lack of data about
actual user behavior.
In the context of designing a Social Cloud, there are various apparent use cases
how simulation can benefit the development process. In particular, the SES can be
used to study following sample questions:
Incentive Schemes Do certain incentive schemes achieve more sharing activities?
What attributes of an incentive scheme are most useful for certain user types?
How does an incentive scheme have to be adjusted to reflect a change of user
types in the system?
User Types and Network Structure What is the effect of user availability on system performance? How is the exchange affected by different utility function
distributions of users (e.g., having more selfish users)? How does the network
structure (e.g., random versus small-world structures) affect collaboration?
How many links between users have to exist before the exchange activities
are effective?
Co-operative Infrastructures Is it feasible to let the users of a Social Cloud platform provide its infrastructure? What is the influence of the user characteristics, such as their availability or inherent willingness to share, on the feasibility of this approach?

34

Resource Sharing in Social Contexts

Resource Exchange and User Strategies How do users select if and how much
they exchange? What effects have different exchange mechanism on the resource allocation? What is the effect of different learning strategies on the
resulting exchange behavior?
Resource Allocation Mechanisms Given a certain supply and demand of resources, which allocation mechanisms yield the best outcome considering certain market objectives? What are the effects of different preference structures
on the performance of the considered algorithms?
In particular, several of the mentioned scenarios are also applied in this thesis.
Chapter 4 presents case studies concerning the design of incentive schemes and
co-operative infrastructures, and Chapters 5 and 6 apply the SES in the context of
allocation mechanisms.

2.2.2. Requirements and Related Tools
The incorporation of aspects such as social network connections and relationship
types into resource exchange simulation yields specific requirements that the SES
has to address. Therefore, this section first defines the requirements for such a
scenario. The second part of the section compares relevant existing simulation tools
and identifies why it is necessary to create a new simulator that is suitable for social
resource sharing settings.

Requirements for the Social Exchange Simulator
In order to facilitate the engineering process of Social Clouds, the SES has to be
able to correctly model and adequately represent such a system. This involves the
representation of the users, their incentives, behavior, and relationships to other
users; the underlying network with its (potentially dynamic) topology; and the
interaction and resource exchange mechanisms that exist within the system.
Flexible User Model The scenarios of social resource sharing can be diverse, and
this needs to be reflected by the user model. In some cases, a user model
might include specific incentives for certain actions, in other cases characteristics such as user availability might be an important design aspect of the
simulation. Hence, rather than predefining a certain type of user model, the

2.2. SIMULATING SOCIAL CLOUDS

35

SES should allow for a flexible specification and at the same time provide
the necessary classes to facilitate the modeling of various user characteristics
(such as availability).
Integration of Social Network Topologies The topology of an underlying social
network, and especially the position that a user has in the network, can have
important consequences on the behavior of both the system and the user. For
example, the importance of a user in a network can depend on the respective
position, such as being a link between different groups that enables the dissemination of information. Another example is the implementation of different feedback mechanisms, which can also be affected by the specific network
structure. For these reasons, the SES has to be able to model different network
topologies (such as random, small world, etc.), and compute certain metrics
based on the user position in the network.
Representation of Relationship Types As discussed earlier, relationships between users can be of various types, and the types might affect the interaction
between them (and thereby affect the entire system). The relationship type,
for example, can affect the respective preference for sharing specific resources
with another person, or can influence the level of trust towards that person.
Therefore, the SES should provide the means to model various types of relationships to study how specific types (or distribution of types) can affect the
interactions between users.
Implementation of Different Resource Exchange Mechanisms The exchange of
resources between users can be facilitated by many different mechanisms.
The simulator has to either model or provide the means to incorporate different types of resource exchange mechanisms and allow for a seamless integration and switch of the used exchange mechanism. In particular, due to the
focus on preference-based matching, it has to provide the means to compute
allocations based on two-sided matching algorithms.

Comparison to Other Simulation Tools
The design and development of simulators as general-purpose tools has been
prevalent in many adjunct areas. This section briefly outlines the most relevant
work for the SES in general.

36

Resource Sharing in Social Contexts

Simulation tools can be broadly divided into two categories:
1. Multi-purpose simulation tools that provide the tools and means to implement
custom simulations.
2. Specialized tools that focus on certain use cases and provide specialized functionalities for these use cases.
Table 2.1.: Comparison of Simulation Tools
Tool

Flexible User Model

Social Network
Integration

Relationship Type
Specification

Resource Exchange
Mechanisms

Multi-Purpose Simulation Tools
Repast
NetLogo
MASON
Specialized Simulators
GridSim
SimGrid
CloudSim
PeerSim
OverSim
SIGVerse
This Approach
Haas et al. (2012)

Table 2.1 shows an overview of the related simulation tools according to their category, and the degree to which they fulfill the previously introduced requirements.
The tools of the first category all belong to the group of general-purpose simulation
tools. Due to having the best match with respect to the mentioned requirements,
the tools RepastSimphony6 , NetLogo7 , and MASON8 are selected for comparison.
These tools represent a system as collection of (often intelligent) agents with certain
policies, and the system behavior is determined by the interactions between these
agents. Most related simulation tools provide functionality for modeling networks,
graphs, and interactions between users. The user model is usually quite flexible
and can be extended by adding custom extensions. However, whereas RepastSimphony, NetLogo, and MASON provide built-in social network functionalities
(which are, however, restricted to the provided libraries and not easily extendable),
the inclusion of relationship types for interactions is only partially supported and
would need extensive customization. Furthermore, none of the three mentioned

6 http://repast.sourceforge.net/

– last accessed May 2014
– last accessed May 2014
8 http://cs.gmu.edu/~eclab/projects/mason/ – last accessed May 2014
7 http://ccl.northwestern.edu/netlogo/

2.2. SIMULATING SOCIAL CLOUDS

37

simulation tools has built-in support for resource allocation mechanisms, which
are a central requirement.
The second category includes specialized simulators such as GridSim (Buyya and
Murshed, 2002), SimGrid (Casanova et al., 2008), and CloudSim (Calheiros et al.,
2011) for the study of complex Grid and Cloud Systems, respectively. These simulators provide tools specifically catered for the needs to represent complex Grid
and Cloud system, including resource allocation mechanisms. The extension of the
user model would involve extensive customization in both cases, as this was not
the primary purpose of these tools. Unfortunately, neither of these three tools provides support for including social network aspects or the integration of relationship
types between users. PeerSim (Montresor and Jelasity, 2009) focuses on simulating
(large) P2P systems. It provides modules to define the underlying network topology, yet the user model is not very flexible and resource exchange mechanisms are
missing entirely. From a more technical point of view, OverSim (Baumgart et al.,
2007) is able to simulate P2P overlay networks, and was developed for the simulation of overlay protocols. Although it provides functionality to model different
types of underlying networks, due to the more technical use case it is not very
suitable for simulating social exchanges. Furthermore, the tool SIGVerse (Inamura
et al., 2010) was developed to model social interactions between humans. However, its focus is more on simulating direct interactions between single humans (or
artificial agents), and does not consider resource exchange or underlying network
connections.
From a comparison of the technical specifications of the mentioned simulation
tools, as presented in Table 2.1, it is apparent that no single simulation tool fulfills
all the necessary requirements for the given scenario. Based on this observation,
two approaches can be pursued: either the extension and customization of an existing tool, or building a separate simulation tool. As the functionalities provided by
existing tools is mostly limited to the included libraries and customization would
have to take the dependencies between new and existing libraries into account, the
design of a separate simulation tool was considered to be the best decision.
As an agent-based representation is a very natural approach to model social systems, where each user (agent) is assumed to have a certain level of intelligence, the
principle of agent-based simulation is adopted for the SES. Using existing libraries
for social network integration, it provides its own libraries to model different relationship types and resource exchange mechanisms, and is thus suited for ad-

38

Resource Sharing in Social Contexts

dressing the design challenges for a simulator for social resource sharing. The next
section provides details about the architecture and implementation of the SES.

2.2.3. Architecture
Social ExchangeSimulator
User
Applications

NetworkTopology

Exchange
Mechanisms

IncentiveScheme

Mechanism

Currency

Monitoring

Experiment
Controller

Artifact

…

…

Application Sensor

ExchangeSensor

Runtime

Core
Elements

UserSensor
User

Resource

Relationship

TrustContext
Resource Sensor

Figure 2.4.: Layered Structure of the Social Exchange Simulator

Similar to comparable simulation tools (e.g. Buyya and Murshed (2002); Calheiros
et al. (2011)) the SES is built through a layered structure, as this permits the flexible
usage and exchange of single components. The implementation language is Java.
Figure 2.4 shows the package structure of the SES. The architecture distinguishes
between three layers: the component layer, the exchange layer, and the application
layer. In addition, the SES contains monitoring sensors in each layer to observe
the state of the exchange network as a whole and its entities, e.g. users, resources,
collaboration and exchange mechanisms, etc. Sensors can be triggered by events,
such as a transaction on the market.
As a foundation, the component layer contains the core elements that are common
to all social resource sharing platforms and necessary for their representation. Examples include the representation of users, resources, different relationship types
and the ability to model trust between users. This can be, for example, implemented through a context-specific representation, i.e. considering the form of exchange as well as the real world relationship between participating users. The
component layer addresses the requirements for a flexible user model as well as
the specification of different relationship types. Above the component layer, the
exchange layer targets the requirement for resource exchange mechanisms and defines how the exchange of resources in the platform takes place. Examples include
the Mechanism for exchange (e.g., auctions, reciprocal exchanges, preference-based

2.2. SIMULATING SOCIAL CLOUDS

39

matching, donations, and forms of volunteer computing), the Artifacts that represent the exchanged resources (single resources, resource bundles, etc.), and the
Currency of exchange (e.g. virtual/sharing credits, preferences, real world currencies, tokens, etc.). Along with a basic Runtime utility class, which manages the
component and the exchange layers, these two layers form the base packages of
the SES. On top of this base, an application layer contains classes to aid the implementation of applications and scenarios. This includes the implementation of different network types, such as small-world or random networks, in order to address
the corresponding requirement. The code base is also not specific for a simulation
environment, and has been designed such that it could also support a real world
implementation of a Social Cloud through, for example, the inclusion of plugins
for social network platforms.
Due to the hierarchical approach, all core and exchange components are designed
to be interchangeable in a plug-and-play fashion. This is done via concrete interface definitions (specified in an API) and dynamic class loading of interface implementations by the Runtime. By doing this, users of the tool can flexibly exchange
classes to study their dependencies and influences on the system. Key examples
here are Mechanisms, which have a crucial impact on the resource exchange, and
Users, who define whether exchange will take place and in what forms. However,
the general interfaces permit the definition of any Mechanism or User type (e.g.
free-rider or altruist, different utility functions, etc.), which are then managed by
the Runtime as plug-and-play components.
Throughout the thesis, the simulation tool will be used to address research questions 1.2, 1.3, 2.1, 2.2, and 2.3 stated in Chapter 1. For example, in Chapter 4 the
SES is used to study the effects of an incentive scheme on a given system, as well
as on different user types. In the second part of the chapter, the feasibility of a
co-operative infrastructure approach is studied through the SES. Finally, Chapters
5 and 6 use the SES to compare different algorithms for preference-based resource
allocation, their dependency on certain factors and the effects of preference manipulation on the matching outcome.

40

Resource Sharing in Social Contexts

2.3. Coordination Challenges in Social Clouds
Common to other sharing platforms, Social Clouds involve certain challenges that
system designers have to be aware of and that need to be addressed to create a
sustainable and successful sharing platform. Specifically, this thesis focuses on two
key challenges: the incentivization of users, and the market-based allocation of
resources. This section, therefore, introduces and discusses these challenges.

2.3.1. Participation Incentives
The task of identifying and engineering incentives for online-based collaboration
and exchange systems is challenging and requires a thorough understanding of
the related concepts. Active participation is crucial for the survival of a resource
exchange platform, as computer-mediated communities (such as Social Clouds)
require a certain minimum number of users (critical mass) for sustained activity
(Markus, 1987). In contrast to anonymous platforms, Social Clouds can leverage
the connections between users to provide incentives for continued participation.
To design participation incentives, it is necessary to understand the motivations
for users to participate in the considered system. Motivation theories have been
developed that aim at defining, categorizing, and explaining the types of motivations that act as underlying drivers of human behavior in certain situations. This
understanding of underlying motivations is crucial for the design of incentives,
as participants might have different motivations in different systems. For example, participants might be motivated through monetary payoffs in some systems,
whereas more altruistically motivated in other systems, in which case an incentive
scheme should concentrate on monetary incentive in the first, and more intrinsic
(non-monetary) incentives (such as social comparisons) in the latter case. Thus,
this section starts with an overview and categorization of motivation and incentive
types. Subsequently, incentive issues in the context of different scientific disciplines
are discussed.

Types of Motivation and Incentives
One of the most common categorizations of motivation types differentiates between the reasons that lead to a certain behavior or action. A prominent example,

2.3. COORDINATION CHALLENGES IN SOCIAL CLOUDS

41

which has been used in a variety of contexts and will also applied in this work,
is the classification of intrinsic motivation and extrinsic motivation. According to
Deci and Ryan (1985), these two concepts can be described as follows:
• Intrinsic Motivation for a certain task means that users find the task itself
enjoyable or interesting (often lacking a directly observable, or separable, reward).
• Extrinsic Motivation stems from situations where the reason for a certain
action is based on a separable outcome, e.g. monetary gain.
Although this coarse distinction of intrinsic and extrinsic motivation is broadly
used in the literature, it is often extended to include sub-types of motivation within
the two categories. In contrast to the distinction of extrinsic and intrinsic motivations, the classification of sub-types of motivation and incentives into the two categories is not as clear and still subject to debate. Additionally, the meaning of different sub-types might differ between different disciplines and authors. For example,
Ryan and Deci (2000) distinguish between four different types of extrinsic motivations depending on the “perceived locus of causality”, i.e., based on the perception
of where the motivation emanates from. To provide an overview of the motivation
types that are most relevant for this work, Table 2.2 shows several sub-types and
the their explanations.9
Table 2.2.: Overview of Incentive Classification and Types
Classification

Intrinsic

Incentive Type

Sample Authors

Altruism
Fun, Enjoyment
Ideology

Clary et al. (1998)
Deci and Ryan (1985); Nov (2007)
Hars and Ou (2001); Lakhani and Wolf (2005);
Nov (2007)
Ryan and Deci (2000); Bishop (2007)
Chan et al. (2004); Nov (2007)
Clary et al. (1998); Hars and Ou (2001)

Interest, Curiosity
Exchange of Knowledge and Experience
Understanding, Learning
Reciprocity
Community Identity
Reputation, Feedback
Extrinsic

Finding of and Interaction with Friends

Kollock (1999)
Hars and Ou (2001); Lakhani and Wolf (2005)
Kollock (1999); Hars and Ou (2001); Rafaeli
and Ariel (2008)
Antikainen (2011)

Own Demand, Influence on Development/Growth
Human Capital, Self-Marketing
Monetary Reward

Hars and Ou (2001); Lakhani and Wolf (2005)
Hars and Ou (2001); Lakhani and Wolf (2005)
Hars and Ou (2001); Tedjamulia et al. (2005)

Within the intrinsic motivation types, altruism is one of the most commonly mentioned types. In economics, altruism refers to “costly acts that confer economic
9 This

classification is not exhaustive, yet provides an overview of the most relevant motivation
types for this work.

42

Resource Sharing in Social Contexts

benefits on other individuals” (Fehr and Fischbacher, 2003, p.785). An example
of this motivation type is the provisioning of resources for volunteer computing
projects, in which case the provisioning infers (energy and opportunity) costs for
the providing users. Besides altruism, fun and enjoyment as well as ideology are
often mentioned, for example in the case of the knowledge-sharing community
Wikipedia (Nov, 2007). Other types of intrinsic motivation also include interest
and curiosity in the activity itself, the intention to share and exchange experience,
and the learning of new skills and knowledge (Hars and Ou, 2001).
Considering extrinsic motivation types, Table 2.2 splits this category into two subgroups. The first group (sometimes considered a distinct motivation type besides
intrinsic and extrinsic: social motivation), subsumes motivations stemming from interactions with other users. This includes reputation-based incentives to increase
one’s standing within a community, the finding of and interaction with friends,
and the identification with a particular community. Reciprocation also falls in this
category and refers to situations where participants either take a certain action as
response to a previous action of other users (such as lending a resource after borrowing another resource from this user), or in anticipation of actions in the future
(considering expected behavior). The second group focuses on extrinsic motivations which are not directly community-based. This involves participating because
one needs or benefits from the project (for example in case of open source projects),
the marketing of one’s skills for career reasons, the increase in human capital, and
also monetary rewards from participation.
Considering the interplay of intrinsic and extrinsic motivations, there seems to be
a complex and context-dependent relationship between the two concepts which is
not necessarily linear. Having both types of incentives in a system, or adding additional incentives, does not have to result in an increase in overall motivation, but
rather to a reassessment of the relative importance of the incentive types (Frey and
Jegen, 2001; Bénabou and Tirole, 2003). A prominent example of this interplay is
the motivation crowding(-out) effect, which shows that adding extrinsic incentives
(such as money) can as well undermine the relevance of intrinsic incentives and
thus affect the actions in an unexpected way (e.g., reducing overall participation,
see Frey and Jegen (2001)). In the context of donating to charities, Ariely et al. (2009)
study the interaction of intrinsic, extrinsic and “image motivation” (similar to social motivations), and focus especially on the interaction of image motivation and
extrinsic motivation. They show that the relative importance of image motivation
depends on the visibility of the specific action, and is higher if other users are able

2.3. COORDINATION CHALLENGES IN SOCIAL CLOUDS

43

to perceive the actions of others. Furthermore, Tedjamulia et al. (2005) suggest a
framework for user participation that considers both intrinsic and extrinsic incentives, yet do not provide an empirical evaluation of their model. Overall, however,
the interplay of different types of motivations is still subject of further research.
After the introduction of different motivation categories and types, approaches and
theories from different scientific disciplines as well as their specific focus and assumptions are discussed.

Motivation and Incentive Theories in Economics, Sociology, and Psychology
User participation is vital for many economic and social systems, and the incentivization to induce desired or proper participation and behavior has been a main
research focus over a substantial period of time. Researchers from different fields
and disciplines have developed models of motivation and participation theory that
try to classify and explain the different types of motivations of users. Two views
that are directly relevant to this work are economics and (social) psychology, which
try to describe and explain user participation and behavior based on the respective
theories.

The Economic View: From Traditional to Behavioral Economics
In traditional economic theory, users are modeled as fully rational agents that maximize their expected payoff/utility in a given scenario (homo economicus, Henrich
et al. (2001)). If rational users engage in a certain (economic) mechanism, e.g. an
electronic marketplace, in order to achieve a desired user behavior the mechanism
has to provide appropriate incentives to the rational users. This is the subject of
the economic area of Mechanism Design10 which aims at designing mechanisms
in a way that prescribes a certain behavior for rational users (Nisan and Ronen,
1999). For the modeling of such a mechanism, users are commonly assumed to
have certain private information, e.g. their valuations for outcomes of the mechanism, and a certain range of actions they can take. They are able to formulate
strategies about which action to take depending on the current state of the system.
Depending on the mechanism, it might not be best for users to reveal this private
information truthfully to the system. For example, they might not provide true
information about their valuations for a certain outcome of the mechanism. An
10 A

formal notation of Mechanism Design is provided in Section 2.3.2.

44

Resource Sharing in Social Contexts

important concept, closely related to the design of incentives, is incentive compatibility, which is achieved if it is the best strategy of a user to reveal their valuations
truthfully. Hence, in Mechanism Design incentives are provided through rules of
the mechanism and the strategies that these rules prescribe on the (rational) user.
There are various examples of the application of a Mechanism Design approach to
provide certain incentives for participating users. For the design of online resource
exchange platforms, it has been applied in the design of (virtual) credit- or moneybased electronic exchange systems or marketplaces (Golle et al., 2001; Antoniadis,
2004; Anagnostakis and Greenwald, 2004; Feldman et al., 2004; Ranganathan et al.,
2004). For example, Golle et al. (2001) introduce an exchange mechanism that uses
micro-payments to incentivize users to contribute in the P2P network. As a certain amount of credits is needed to consume resources and credits can be gained
by providing resources, such an incentive scheme aims to encourage participation
and resource sharing while at the same time preventing users who only consume
without contributing themselves (free-riders) from exploiting the system.
While such approaches fit nicely in the traditional economic theory and the corresponding mechanisms have certain (theoretical) properties, over the years it has
been shown that there are several issues with these types of incentive schemes.
On the one hand, despite their nice theoretical properties there seem to be several reasons why especially (micro-) payment-based schemes are not well accepted
(Odlyzko, 2003). On the other hand, there is substantial critique on the perfectly rational model of users. This is further exemplified by various experiments that show
that users do not act purely based on monetary preferences as usually assumed in
traditional economics (see e.g. user behavior in economic games: Andreoni and
Miller (2002); Fehr and Schmidt (2006)).
In contrast to pure self-utility maximization, users might also be contributing to a
platform for other reasons, such as social or ideological reasons (Mowbray et al.,
2006). The field of Behavioral Economics aims at explaining the discrepancy between
assumed and observed behavior of participants in economic systems. Most theories in behavioral economics still use the concept of utility-maximizing users,
yet acknowledge that non-monetary incentives might influence and explain the
observed behavior. The most common approach is to augment the utility functions with additional factors that model non-monetary concepts. For example, (expected) reciprocity tries to explain how individuals incorporate the expected return
from other users in their decision process, and inequality aversion assumes that users

2.3. COORDINATION CHALLENGES IN SOCIAL CLOUDS

45

want to avoid highly unequal payoffs. Examples for these theories are the theory
of equity, reciprocity and competition (Bolton and Ockenfels, 2000), the theory of
fairness and reciprocity (Fehr and Schmidt, 2000, 2006), and the concept of social
preferences (Charness and Rabin, 2000). These theories are successfully able to explain a large variety of observed user behavior in different economic settings, and
acknowledge the fact that different users might put different emphasis on nonmonetary motivations.

User Motivation Theories in Sociology and Psychology
Whereas economic theory commonly considers incentives based on the assumption
of utility-maximizing (rational) users, over the past decades (social) psychology has
developed its own set of different motivation theories. Although a comprehensive
review of motivation theories is outside the scope of this section, some theories
relevant for the problem at hand will be introduced.
One of the earliest works, which is still influential in incentive and learning theory,
is the work by Thorndike (1927) and Skinner (1953). Thorndike (1927) describes
the “Law of Effect”, which states that given a certain set of potential actions, which
are repeated over time, the actions that are beneficial are chosen proportionally
more often over time, whereas actions that are detrimental are chosen less often.
Skinner (1953) builds on this fundamental result, which is also the building block
of Reinforcement Learning theory, and proposes the “operant conditioning” theory.
It states that actions are motivated by (different types of) rewards, and differentiates whether the motivations are triggered exogenously (extrinsic motivations,
e.g. money) or endogenously (intrinsic motivations, e.g. fun). Closely related is
the “expectancy-value theory” which considers beliefs about abilities, expectancies
for success as well as subjective task values as components determining motivation
(Wigfield and Eccles, 2000). According to Ryan and Deci (2000), research that builds
on the operant theory considers intrinsic motivation as a case where the task is the
reward itself, and related research is mainly interested in studying what makes a
task intrinsically interesting. In contrast, research building on the learning theory,
which considers a needs-based characterization of motivations, aims at studying
which (psychological) needs are addressed by intrinsically motivated tasks.
Beside reward-based theories, another branch of motivation theory considers the
social fabrics involved in the resource exchange setting. The theories of social comparison (see e.g. Festinger (1954) and Suls et al. (2002)) try to explain how the moti-

46

Resource Sharing in Social Contexts

vation of participants depends on their comparison with their peers. The theory of
“social loafing” argues that users within groups tend to work less (have lower motivation to contribute) than if they would work individually (Karau and Williams,
1993). Furthermore, the “Common Identity Theory” and the “Common Bond Theory”, which describe the motivations of individuals based on their connection to
the (social) group, or other members of the (social) group, respectively, have been
applied in the context of designing online communities as well (Ren et al., 2007).
Other theories from sociology include the Resource Theory of Social Exchange (Foa
and Foa, 2012) which provides a framework for analyzing the exchange of several
types of resources, e.g. money or services, in social contexts. Similarly, the Social
Exchange Theory with its focus on organizational behavior considers issues such as
reciprocity and negotiations in social contexts (see e.g. Emerson (1976); Cropanzano and Mitchell (2005)).
Lastly, considering the participation in volunteer scenarios, the volunteering motivations scale of Clary et al. (1998) distinguishes between six functions that help
to explain volunteer behavior in various settings. Given a volunteer task, value is
related to altruistic concerns, understanding addresses the ability to gain new skills
and learning experiences, social considers reputation and social comparison, career targets career-relevant beneficial effects of participating in the task, protective
is ego-related and explains volunteer participation to avoid negative feelings and
guilt, and enhancement focuses on personal development.

Summary
The focus of this section was to provide an overview of concepts and theories relevant for user participation. First, different motivations to participate and contribute
in sharing and exchange scenarios were categorized. Afterwards, relevant theories
from the field of economics, sociology and psychology were discussed.
Overall, it can be seen that the incentivization of user participation and the understanding of underlying motivations is a multi-faceted topic with various theories in
different research fields. In particular, there seems to be a complex interrelationship
between incentives and their relative importance in different scenarios. This also
motivates that the relevant user motivations and incentives have to be identified
for the specific scenario at hand.

2.3. COORDINATION CHALLENGES IN SOCIAL CLOUDS

47

Several of the important concepts described in this section will be used in Chapters
3 and 4, which study user incentives in Social Clouds. In particular, several of the
mentioned motivation types will be used to identify relevant incentives in such a
setting.

2.3.2. Resource Allocation
Mechanism Design
From the perspective of economic theory, the general Social Cloud setting with
users requesting and providing resources is an example of a microeconomic system. Given the economic environment which consists of the participating users, their
available resources and preferences, a humanly-devised institution provides rules
how participants can interact with each other and the institution, and determines
how the exchange of resources is structured. Such an institution prescribes a certain behavior on participants and leads to an economic outcome, e.g., how resources
are allocated. The performance of the considered institution can then be evaluated
through certain system performance metrics (Smith, 1982; Weinhardt et al., 2003).
The previously mentioned task to find allocation mechanisms for Social Clouds
is equivalent to designing an institution that governs this exchange process. The
field of Mechanism Design provides the means to formalize this institution design
process. A mechanism is defined as “specification of a message space for each individual and an outcome function that maps vectors of messages into social decisions
and transfers” (Jackson, 2003, p.2). In the realm of resource exchanges, the message
space defines how the participants interact with the market, e.g., how they specify
their valuation for a certain resource. The outcome function then determines a certain allocation of resources, which is often accompanied with the determination of
certain transfers between participants (e.g., monetary transfers).
Formally, adopting the notation by Jackson (2003), participating users i, i ∈
{1, ..., N } have private information (e.g., about motivations relevant to them). User
i’s private information is represented by its type θi , which is in the set Θi . This type
determines agent i’s preference for a certain outcome o ∈ O, where O is the set of
possible outcomes. The preference (or valuation) is often modeled as utility function, where ui (θi , o ) determines user i‘s utility for a given outcome, given its type

48

Resource Sharing in Social Contexts

θi . For the comparison of two different outcomes, o1 and o2 , ui (θi , o1 ) > ui (θi , o2 )
indicates that outcome o1 is preferred to outcome o2 .
From a system perspective, the allocation of resources would determine a certain
outcome which optimizes a desired goal, e.g., the maximization of total welfare. In
other words, given the (potential) types of users, the system designer can define
a function f : Θ1 × . . . × Θ N → O that finds the best outcome given the user types
(also referred to as Social Choice Function). As the user types are private information, the system designer has to find a way to incentivize users to reveal this
private information in order to be able to find the best outcome. This is the goal of
the mechanism, which defines rules how users can interact with the system. More
specifically, the mechanism prescribes a certain strategy set Si on user i, which represents the actions available to user i. It also defines an outcome rule that maps the
strategy set to a certain outcome. A mechanism is said to implement a social choice
function if, in equilibrium, the mechanism yields the same outcome as the social
choice function. A transfer function t:Θ → Rn determines the transfers that result
from a given outcome. As described subsequently, this transfer can have different
implementations, including monetary and credit transfers as well as the complete
absence of transfers.

Types of Exchange Mechanisms
The allocation of resources is one of the key aspects of market design. Given a
certain supply and demand of resources, an allocation mechanism determines which
request is matched to a given offer and vice versa. There are many examples where
such mechanisms are applied, ranging from technical applications such as job and
Virtual Machine (VM) scheduling on compute infrastructures, to economic applications such as the allocation of products to consumers. The types of allocation
mechanisms that are used in the various settings can be diverse as well. They range
from monetary-based mechanisms like auctions or fixed-price markets to dynamic
negotiation that determines the details of the allocation and exchange of goods and
resources. The most relevant mechanisms can be categorized as follows:
Monetary-based Mechanisms If the exchange and sharing of resources involves
monetary payments, traditional market-based approaches are often used. Examples are fixed-price systems (such as Cloud Computing marketplaces like

2.3. COORDINATION CHALLENGES IN SOCIAL CLOUDS

49

Amazon Web Services11 ) where resources are offered for a specified price,
and auction mechanisms where the price is determined dynamically depending on the current supply and demand (such as auctions on eBay12 ). Such
mechanisms assume that participating users have a certain valuation for the
resources they are interested in, and interact with the mechanism according
to this valuation.
Credit-based Mechanisms Similar to monetary-based mechanisms, platforms
with (virtual) credit-based mechanisms use valuation-based matching procedures. The difference between the two approaches is that credit-based systems do not necessarily involve real money. For example, users can gain
credit by providing resources for other users, and use that credit to consume
resources in return. A key challenge of such credit-based systems is the management of the currency with respect to credit value, specifically inflation and
deflation due to leaving or arriving users (see e.g. Irwin et al. (2005)).
Preference-based Mechanisms In many systems where the use of monetary
mechanisms is not feasible, not wanted, or morally/ethically not possible,
preference-based mechanisms are applied to retain some of the advantages
of a market-based resource matching approach (Roth, 2008). In this case, resources are shared on the basis of preferences that users have for other users,
and the mechanism allocates the resources with respect to certain criteria (e.g.,
overall welfare).
Best-effort and Volunteer Mechanisms In such mechanisms, resources are essentially shared without involving credits or monetary exchanges, or even preferences for other users. Examples are volunteer platforms without direct reward for resource provisioning, and Trophy-based systems in which users
can gain certain trophies for sharing resources (such as the volunteer computing project BOINC, see Anderson (2004)). In contrast to credit-based systems,
however, such trophies or other rewards are not used for the resource exchanges, but rather for the personal gratification and satisfaction of the users.
Decentralized Approaches Another option is to leave resource allocation to the
users themselves, for example via distributed communication protocols (see
e.g. Streitberger and Eymann (2009)). The downside of a decentralized selforganizing approach, however, is that users would most likely have to invest
11 http://aws.amazon.com/
12 http://www.ebay.com

– last accessed May 2014
– last accessed May 2014

50

Resource Sharing in Social Contexts
a significant amount of time in the system trying to get “good” deals, negotiate and in general manage their exchanges. The resulting emergent user behavior would be unpredictable and potentially inefficient at the system level.

Note that the differentiation of mechanism types considers different implementations of transfer functions. In monetary mechanisms, the transfer function specifies
the monetary amount that users have to pay or receive, given the determined allocation. For credit-based systems, instead of monetary transfers the amount of exchanged credits is determined. In case of preference-based mechanisms, there is no
actual transfer of money, credits, or other remuneration. Best-effort and Volunteer
mechanisms also do not involve transfers, yet also do not provide a market-based
type of mechanism. The last example, decentralized allocation approaches, does
not fall into the above categorization as no central mechanism is designed.
To decide which approach is promising in case of a Social Cloud, the advantages
and disadvantages of the approaches can be compared. As mentioned before, the
social context and the complex interactions between monetary and intrinsic or social motivations suggest that non-monetary mechanisms are most suitable. Furthermore, decentralized allocation systems can be problematic in terms of user interaction and feasibility, and can lead to market failure (Roth, 2008). This leaves
centralized, managed market mechanisms as potential options. Considering creditbased systems, their ability to match resources according to the valuations of users
and relative supply and demand stands in contrast to the significant challenges of
managing the credit currency. The potentially high interaction of users with the
system to determine and submit their valuations for resources can also be seen as
a drawback. In contrast, volunteer or best-effort mechanisms which do not encompass common market criteria such as system welfare or fairness might be easier
to manage, yet produce allocations that can be sub-optimal with respect to these
criteria. Finally, preference-based matching can be seen as a compromise between
market efficiency and ease of use, which not only addresses the resource allocation
problem, but also considers the social context of allocations.

Two-Sided Matching for Preference-based Resource Allocation
The field of two-sided matching markets is a successful and established means to
allocate resources based on preferences rather than monetary or credit-based valuations. Instead of monetary valuations, it assumes that users have an ordinal pref-

2.4. SUMMARY

51

erence ranking with whom they want to share, i.e., both sides have preferences for
the other side. Each side ranks participants of the other side in an ordinal ranking
(with rank 1 being the most preferred choice). The objective of two-sided matching
is to guarantee that the solutions given by such a market mechanism satisfy certain
desirable characteristics, such as stability, fairness, or (social) welfare.13
In the context of the previously introduced Mechanism Design notion, a twosided matching mechanism specifies a message space (users submit their preference rankings and get the final match), and the outcome function (matching algorithm) matches the preferences to a certain outcome (specifies which users are
matched). The transfer function, in this case, is not further specified as the mechanism does not use monetary transfers.
Overall, the designer of a Social Cloud platform has the objective to guarantee
certain performance criteria for the exchange. This raises an important question:
Which economic systems can be implemented for social resource sharing such that
certain economic properties are still fulfilled? As the described two-sided matching
approach is in line with the social context of the resource exchange in Social Clouds,
preference-based matching mechanisms will be the focus of Chapters 5 and 6.

2.4. Summary
This section introduced the basic concepts and foundations that are relevant for
the thesis. First, Section 2.1 introduced Social Clouds as the unifying use case that
is applied throughout the thesis. Furthermore, the prototype of a Social Compute
Cloud and details of its technical implementation were presented. Section 2.2 described the simulation tool that can be used as complementary methodology in the
design of social research exchange platforms. The simulation tool will be used for
several evaluations throughout the thesis. Finally, Section 2.3 discussed two relevant coordination challenges in the context of Social Clouds: the incentivization of
user participation as well as the allocation of available resources, which are both
considered in this thesis.
Based on these foundations, Part II considers the first challenge identified in Section 2.3: the identification of relevant user participation incentives and the design of
corresponding incentive schemes. Part III of the thesis concentrates on the second
challenge, resource allocation in Social Clouds through preference-based matching.
13 Exact

definitions of the metrics are provided in Section 5.1.2.

Part II.
Incentive Engineering for Social
Clouds

Chapter 3.
Incentives in Social Clouds
“The proliferation of online communities may suggest that the design of a community for a
particular purpose is straightforward. Unfortunately, this is not the case.”
(Cheng and Vassileva, 2006)

NLINE communities, sharing platforms, and other network-based exchange
systems stand and fall by users actively participating on the platform. This
seemingly trivial observation has been a valuable, sometimes harsh lesson for
many online communities during the past decades. A fitting example is the emergence of various social network platforms and their struggle to achieve a critical
mass of users that ensures the platform’s sustainability (Westland, 2010). Over
the course of several years only few of these platforms survived, with Facebook
emerging as the most prominent site. Others disappeared due to a lack of actively
participating users.

O

In the context of Social Clouds, the same important user incentive challenges remain. As with all market platforms, a resource sharing mechanism such as a Social
Cloud depends on the active participation of its users in creating a sustainable resource exchange. Hence, providing proper incentives for users throughout their
interactions with the platform can be seen as one of the fundamental challenges in
the creation and management of sharing platforms.
Whereas the importance of user participation is generally agreed on and incentive
issues have been studied for a wide variety of platforms and online communities,
the study of user motivation and incentivization in specific scenarios is still an area

55

56

Incentives in Social Clouds

of active research. User motivation can be quite complex and depends on the specific characteristics of the scenario (see e.g. Ryan and Deci (2000)). For example,
the disparity of free-riding behavior in (mostly anonymous) public goods markets
(such as P2P platforms) stands in contrast to the considerable quantity of charitable
and volunteer activities in other settings (Ariely et al., 2009). Furthermore, different participation activities such as passively consuming information compared to
active posting of information, are influenced by different factors and thus require
specific stimuli (Koh et al., 2007).
Section 2.3.1 categorized motivation into intrinsic and extrinsic types and discussed
that their interplay is far from trivial. In particular, the design of incentives within a
particular system might have unexpected consequences on user participation (Frey
and Jegen, 2001; Bénabou and Tirole, 2003). It is therefore necessary to consider the
motivation and incentive issues in a platform-specific way through the identification and implementation of incentives relevant for the considered scenario. This
chapter introduces and analyzes incentive challenges in Social Clouds. Its goal is
to address research question 1.1 as stated in Chapter 1.2.
R ESEARCH Q UESTION 1.1 ≺ I NCENTIVE E NGINEERING  What are the stages of
participation and the corresponding relevant incentives that users exhibit in Social Clouds?
For the design of an incentive scheme, several steps must be taken. In the general context of online communities, several distinct user participation stages have
been previously identified (Jones and Rafaeli, 1999; Iriberri and Leroy, 2009). These
stages define how users interact with the community or platform. On social resource sharing platforms such as a Social Cloud, the interaction of users with the
platform seems to consist of different stages of participation. Starting with the discovery and registration on the platform, participation can evolve into active contribution and sharing of resources. Hence, the first step is the identification of different stages of participation during which a user interacts with and participates in a
Social Cloud. This is the focus of the first part of this chapter. This understanding of
participation stages then helps in the identification of relevant incentives for each
stage, as the importance of certain incentives may change between stages. For example, from a platform perspective the signing-up stage has a different goal than
the active participation stage, and different incentives might be necessary (Fogg
and Eckles, 2007).

3.1. PARTICIPATION INCENTIVES IN RESOURCE SHARING PLATFORMS 57
After identifying the relevant participation stages with their characteristics and
challenges, it is necessary to appropriately design and engineer specific incentives
to stimulate user participation in each of these stages. Such a process involves two
steps: the identification of relevant incentives in the particular stages, as well as the
discovery of factors that influence the relative importance of said incentives. This
is addressed in the second part of this chapter.
Potential motivations and incentives have been discussed in Section 2.3.1. Among
the factors that can have an influence on the perceived relevance of certain incentives for different users are user characteristics (such as personality profiles), the
type of resource that is exchanged, and the setting of the exchange (e.g., sharing
between friends vs. sharing with acquaintances).
After an introduction of important concepts and related work in Section 3.1, Section 3.2 identifies the different participation stages of user interaction with a Social Cloud and provides a discussion of their respective characteristics and challenges. In Section 3.3, the results of a small-scale web-based survey are presented
to demonstrate how relevant user incentives in the different participation stages
can be identified. Finally, Section 3.4 concludes this chapter with a discussion of
the findings and an outlook on future work in this area.

3.1. Participation Incentives in Resource Sharing
Platforms
As the need for proper incentivization is universally agreed upon (see Section 2.3.1)
this section extends the incentive concepts introduced in Chapter 2 by providing
an overview of the most relevant related work in the field of online and virtual
communities, as well as similar resource sharing systems. First, due to its close
conceptual background, Section 3.1.1 discusses incentives to participate in online
communities in general. Second, Section 3.1.2 summarizes incentivization problems in similar systems, among which P2P sharing networks are a prominent example.

58

Incentives in Social Clouds

3.1.1. Incentives in Online Communities
Social resource sharing platforms such as Social Clouds are a special form of online
(virtual) community. Starting with the work of Rheingold (1993), online communities have been considered as a type of community that uses and leverages computer
technology for communication among members of the community. The original
definition is as follows:1
Definition 4 ( Virtual Community, Rheingold 1993 ). Virtual communities are social
aggregations that emerge from the Net when enough people carry on those public discussions long enough, with sufficient human feeling, to form webs of personal relationships in
cyberspace.
Once online communities grew larger and more important over the years, researchers became more interested in various concepts revolving around these communities. Hence, a review of related literature is helpful in discussing the concepts
that have been discovered and studied since the beginning of online communities.
Among the concepts being discussed here are perspectives on participation from
different scientific disciplines, aspects such as usability, as well as dependencies
between incentives, participation stages, and user types.

Online Communities from Different Research Directions From an economic
viewpoint, the emergence of cooperation among community members has been
studied using concepts such as public and digital goods. For example, Kollock
(1999) applies this approach and further discusses several potential motivations
how the cooperation and contribution of members can be explained, mentioning
aspects such as reciprocity, effects on reputation, and including the community
good in the utility functions of its members. From a business perspective, Williams
and Cothrel (2000) argue that the creation and management of virtual communities is paramount for successful online businesses. They present several examples
of successful online communities and discuss aspects that have been found to be
critical for the studied communities. Two important aspects are reaching a critical
mass of users and the design of appropriate (communication) tools for the community, which are used to gain feedback of community members and use it for the
1 The

terms online community and virtual community are used interchangeably in this chapter.

3.1. PARTICIPATION INCENTIVES IN RESOURCE SHARING PLATFORMS 59
development of the community. Another frequently observed aspect is the recognition of different types of user motivation for participation (see Section 2.3.1 for
an overview), and the existence of different user types with respect to their contribution (Bishop, 2007; van Dijck, 2009).
Besides economic and business perspectives, in recent years there is an increased
effort in trying to understand the social background of online communities. On
one hand, researchers are interested in how the ability to interact and socialize
with other community members affects the community itself (Preece and MaloneyKrichmar, 2003; Lazar and Preece, 2002). On the other hand, from a viewpoint of
social capital theory the effects of underlying social networks on the community
(considering interaction between and benefits for members) have also been studied. For example, Ganley and Lampe (2009) find that the structural properties of
the social network can impact the perceived benefits for members. This is an interesting implication for Social Clouds, as the position of a user within the social
network might affect the decision to join and participate on the platform.

Usability From a technical perspective, usability has also been identified as
important concept related to participation in online communities (Preece and
Maloney-Krichmar, 2003). Considering the relationship of motivation and usability, Wang et al. (2012) apply a technology acceptance model and structural equation
modeling to show that intrinsic motivation significantly influences the perceived
usefulness, the perceived ease of use, and the actual use of the online community.
Furthermore, guidelines and patterns have been developed that try to encourage
user participation and interaction with the platform (see e.g. Porter (2010)). However, Vassileva (2012) argues that these best practices are gained in retrospect from
already successful platforms, whereas generalized recipes are not readily available
if a platform has to be developed from scratch.

Participation Incentives Considering the lifecycle of online communities and
user participation therein, it is also recognized that participation can be structured
in separate phases depending on the lifecycle stage in which the community is.
Fogg and Eckles (2007) present a model for a “behavior chain” for user participation. They distinguish between three phases: discovery, superficial involvement,
and true commitment. They illustrate their model in the context of different web
services and discuss implications for the designer of such a service. Users (or

60

Incentives in Social Clouds

user groups) might also require different incentives depending on the participation stage that they are in (Jones and Rafaeli, 1999). In other words, incentives for
signing-up and initial participation on a platform might be different than incentives for continued (long-term) participation. Cheng and Vassileva (2006) discuss
that incentives are needed to get a critical mass of participants, yet argue that too
many low-quality contributions (“information overload”) can also be detrimental
to the community as some users might be more inclined to leave in such a case.
Therefore, they argue for a balance of quality and quantity for participating in a
resource-sharing community, and propose an adaptive incentive mechanism that
takes into account both the user profile (such as their reputation) as well as the
current needs of the community.
Given that literature identifies several different participation incentives, it is generally agreed that different users perceive these incentives differently. Hence, this
calls for a personalization of incentives as incentive schemes which do not distinguish between different user types or user behavior lack the capability of addressing this fact. The theory of “User and Group (Community) Modeling” is an
example how this can be addressed (Vassileva, 2012). Furthermore, Lampe et al.
(2010) apply the “Uses and Gratification” theory as well as “Organizational Commitment” theory, and find that a “feeling of belonging” to an online community is
a major motivation factor across user types, and that users might continue participating on the platform for reasons other than the ones they had when they initially
joined the community.
Other aspects that have been studied related to participation in online communities
are different types of recognition (for users) and their effect on participation (Chan
et al., 2004), the effects of social comparison on user contribution in online communities (Harper et al., 2007), and the application of gamification aspects in online
communities in order to further foster participation and contribution (Deterding
et al., 2011).

3.1.2. Participation Incentives in Similar Sharing Systems
Social Clouds are conceptually close to other sharing platforms, thus it is helpful
to consider participation incentives in these similar systems. Examples for systems
where the engineering of proper (user) incentives has been studied include file
sharing P2P networks (Golle et al., 2001; Feldman et al., 2004; Anagnostakis and

3.1. PARTICIPATION INCENTIVES IN RESOURCE SHARING PLATFORMS 61
Greenwald, 2004; Antoniadis, 2004; Zhang et al., 2009), Volunteer Computing (Nov
et al., 2010), sharing of workflows in social contexts (De Roure et al., 2009), and
even in technical areas such as traffic regulation in wireless access networks (Liao
et al., 2002).

P2P Networks In contrast to a social scenario where there are direct or indirect
relationships between users, P2P systems are mostly anonymous. Certain characteristics of P2P, such as this anonymity and the possibility of whitewashing, i.e. the
creation of new identities, make appropriate incentivization techniques necessary.
One of the most severe problems in P2P networks is free-riding, i.e. the consumption of resources without providing resources in return. For example, Cuevas et al.
(2010) find that only a small fraction of P2P users create most of its content. Various
approaches have been proposed in this setting, ranging from micro-payments for
contribution (Golle et al., 2001), discriminating server selection and shared history
of exchanges/behavior (Feldman et al., 2004), exchange-based incentives through
service prioritization (Anagnostakis and Greenwald, 2004), to the use of tokens,
reputation and service classes for the exchange of resources (Ranganathan et al.,
2004). Novel incentive schemes continue being published, indicating the practical
importance of this research.
Interestingly, there is evidence that intrinsic motivations are more important than
extrinsic ones which, though sometimes relevant as well, are not considered the
dominant motivation (see e.g. Tedjamulia et al. (2005)). Despite this finding, many
P2P incentive schemes focus on credit- or payment based systems which can be
considered extrinsic incentives. One potential explanation for this is the fact that
the P2P networks considered for these schemes are anonymous, which makes it
potentially riskier to rely on pro-social behavior or generalized reciprocity.

Other Community-based Networks Conceptually close to the use case Social Cloud are the previously mentioned platforms myExperiment.org and
nanoHUB.org. myExperiment.org (De Roure et al., 2009) is an example of workflow
sharing within a scientific community. Here, a platform is provided with which researchers can share, adapt and use scientific workflow processes and experiments,
therefore enabling collaboration and the spread of common workflows. Due to its
target group, the main incentive for scientists to join and participate in the sharing of workflows is the potential improvement of scientific processes through the

62

Incentives in Social Clouds

use of community-fostered workflows and experiments. Similarly, nanoHUB.org
(Klimeck et al., 2008) allows for the sharing of teaching and research materials on
nanotechnology and aims at fostering collaboration in this field. nanoHUB.org
uses virtual credits (“nanos”) to reward contributions such as answering questions
in forums. The amount of credit indicates the level of contribution of a user, and
can also be used to purchase several items in the platform store. Examples of nonscientific communities are online photo-sharing communities, where different motivations such as enjoyment or self-development are found to be important (Nov
et al., 2010), and social commerce communities where users are able to open virtual
stores and are connected to other sellers via a social network (Stephen and Toubia,
2010). In the latter case, monetary rewards are the primary incentive to participate
in such a community.

3.1.3. Discussion
Summarizing the related work, there are two challenges in the design of participation incentives that need to be addressed: 1) the existence of distinct participation
stages with potentially different incentivization requirements, and 2) the acknowledgment that not all users react similarly to certain incentives, which has to be
considered in the design and evaluation of incentive schemes.
Despite the described literature on similar systems, the effects of incentive schemes
on social sharing systems require additional research. For the setting of a Social
Cloud (or a general social resource exchange setting), the different participation
stages (such as discovery followed by active participation) as well as the relevant
incentives therein have to be identified. This is the focus of Section 3.2. In addition, to provide a more general framework for participation stages in social sharing platforms, factors that influence the importance of certain incentives need to
be analyzed. For example, resource sharing in a professional setting (i.e., with colleagues, companies, etc.) might emphasize monetary incentives to a higher degree
than sharing with close friends. The survey discussed in Section 3.3 aims to address
these issues. Finally, considering that user types can have different motivations to
participate, the effect of user type distributions on the effectiveness of an incentive
scheme needs to be studied as well. This is subject of the case studies in Chapter 4,
which aim at taking a first step in this direction.

3.2. ENGINEERING INCENTIVES FOR SOCIAL CLOUDS

63

3.2. Engineering Incentives for Social Clouds
Among the first steps in the design process of incentive schemes is the identification of the different stages of participation that users have with the platform. Only
with this insight, and the incorporation of the characteristics of the participation
stages, the incentive scheme can be tailored to the specific needs of the exchange
platform. This section, therefore, provides a classification of three different participation stages for Social Clouds, as well as a discussion about factors which influence participation within these stages. This section is an extended version of Haas
et al. (2011).

3.2.1. Incentives During the Participation Lifecycle
The sustainability of resource sharing mechanisms crucially depends on having a
critical mass of active users with continued participation over time. Therefore, in
the design of a Social Cloud, providing appropriate incentives for active participation has to be among the most important goals. In order to provide appropriate
incentives, different stages of user participation and contribution can be distinguished (Jones and Rafaeli, 1999; Iriberri and Leroy, 2009). The different stages
of a community life-cycle have different goals, e.g., starting with aim to get a certain number of registered users and subsequently focusing on encouraging useful
contributions. Hence, there is a changing emphasis within these stages on which
aspects successful communities and sharing platforms must focus on (Iriberri and
Leroy, 2009). An effective incentive scheme has to address these different stages
and be adaptable for changing requirements (Vassileva, 2012).
The incentive classification scheme presented in this chapter identifies three participation stages during which users interact with the Social Cloud in different ways.
This is similar to related classifications of user participation phases such as (Fogg
and Eckles, 2007), yet focuses more on the underlying social connections of users.
Each stage can be characterized by unique goals and therefore has specific incentive requirements. Figure 3.1 shows the three stages along with the challenges that
have to be met. In the first stage, User Discovery and Registration, potential users
of the platform have to be discovered and the value of participating on the platform
has to be communicated to them. The next stage, User Participation, addresses the
challenge of incentivizing registered users to actually offer resources, while at the

64

Incentives in Social Clouds

same time discouraging free-riding behavior. The last stage, Social Behavior in
Resource Sharing, is closely intertwined with the second stage, yet focuses on social aspects. If users actively participate and provide resources, they should have
incentives to adhere to certain social behavior. For example, the platform might
implement agreements that specify the responsibilities of the sharing partners in
the resource exchange. A desired behavior, in this case, could be providing the resources according to the agreement rather than defecting from the offer if a request
arrives, or not to engage in malicious behavior such as intentionally providing incorrect feedback about other users (Petri et al., 2012).
User Discovery and
Registration
• Advertisement
• User selection
• Trust aspects

User
Participation
• Incentives for
resource
provisioning
• Free-rider issues

Social Behavior in
Resource Sharing
• Adherence to trading
agreements
• Social behavior

Figure 3.1.: Participation Stages and User Incentivization Problems

Stage 1: User Discovery and Registration
Before users can participate in a resource exchange, they have to be made aware of
its existence and be invited to join the corresponding platform. Here, several steps
can be distinguished, namely discovery, invitation, and registration of potential
users.
For the discovery of users the existing relationships between members of the platform with potential users can be utilized. Both direct discovery through manual
invitations as well as automatic discovery, e.g. through automated advertisements
or scraping of the social networks of users, are feasible. In case of manual discovery, existing users can be incentivized to find new users by, for example, bonus
programs where the user gets a certain amount of credits or other benefits for each
invited user that joins the platform. Gamification aspects can also be utilized for
this task (Deterding et al., 2011), e.g., by providing badges or trophies for successful
user invitations. In case of automatic discovery, mining of available data through
the underlying social network platform can be applied to automatically identify
and suggest users. Similar to the process of user discovery, invitations can then be
sent either manually, or automatically. If potential users are discovered and invited

3.2. ENGINEERING INCENTIVES FOR SOCIAL CLOUDS

65

to join the platform, their decision depends on their motivation and the incentives
to join the platform. Here, too, several incentivization schemes are possible. For
example, invited users who join the platform may receive a sign-up bonus, e.g.
credits, or may be awarded priority functionality during a certain time after their
subscription. The actual form of incentivization will most likely depend on the
specific application scenario.
An important issue that has to be considered is the concept of trust. For example, if the platform (and thereby the social network) gets larger users might not
be direct friends of each other, and with such an increase in size and complexity
new assumptions about trust may be necessary. The problem of trust transitivity
is particularly interesting in this case. Some approaches simply consider the existence of trust transitivity and calculate indirect trust relationships through multiplication of direct relationships (Golbeck, 2005). Transitivity of trust might be,
however, highly user-dependent, making such a general assumption problematic.
Caton et al. (2012) discuss these and other aspects of trust, specifically in the context
of Social Clouds.

Stage 2: Encouraging Active Participation
After the discovery and subscription of new users, they have to be incentivized to
actively participate on the platform and offer resources. Only if a certain, critical
mass of actively participating users is achieved, will the platform be sustainable
(Preece and Maloney-Krichmar, 2003). A lack of participation is considered a danger for such virtual communities (Rafaeli and Ariel, 2008).

Tackling the Free-rider Issue Providing incentives for participation is closely
connected to the free-rider problem. Naturally, users may perceive the consumption
of resources as beneficial whereas the provision of one’s resources usually induces
costs of some sort, e.g. power consumption. It is a well-known problem in online
communities that users may not have adequate incentives to actively participate,
which leads to a small number of contributors and a large number of passive consumers. Whereas the contributors actively engage in the community and make up
much of the overall participation on the platform, the passive users, sometimes referred to as “lurkers”, consume resources without providing resources themselves
(Bishop, 2007; van Dijck, 2009). Example communities where this phenomenon can

66

Incentives in Social Clouds

be observed are P2P networks (Adar and Huberman, 2000), open-source communities (Lakhani and von Hippel, 2003), and the platform Wikipedia (Tapscott and
Williams, 2007; Priedhorsky et al., 2007). Hence, it is necessary to design the platform in such a way that users profit from contributing. In the long run, this ability
will be one of the make-or-break factors of a resource sharing platform.

User Heterogeneity Another issue is the potential heterogeneity of user types
with respect to their motivations to participate and contribute. Previous work (see
e.g. Andreoni and Miller (2002) for user behavior in social settings) has shown that
in exchange scenarios, different user types can be distinguished which have different perceptions and motivations to contribute. This poses a significant challenge in
the design of a resource exchange platform such as a Social Cloud. For the design of
incentives, this has the crucial implication that incentives have to be provided that
address all or most of the potential user types of the community. Ideally, the incentive scheme takes the different user types into account and provides individualized
incentives (Vassileva, 2012). One common form of incentive schemes that try to address both free-riding behavior and heterogeneous user types are schemes where
users obtain participation points or rewards from contribution, and are only allowed to participate in case they can provide enough points or a high-enough level
of previous participation (see e.g. Ranganathan et al. (2004) on P2P systems).
User participation also crucially depends on the chosen economic allocation mechanism. Depending on the type of user and the application scenario, different market mechanisms provide different incentives for active participation. This issue is
elaborated in Section 3.2.2. A case study how heterogeneous user types can be
addressed in the design of an incentive scheme is provided in Chapter 4.2.

Stage 3: Incentivizing Social Behavior
Even if users share and offer different types of resources, it is not guaranteed that
they will adhere to their offers and provide the resources as anticipated. Considering the social setting, social behavior in this case can be defined as the adherence to
the platform rules and agreements, as well as refraining from malicious behavior.
The willingness to provide resources is a necessary first step to get a reasonable
amount of offers as well as heterogeneity that increases the attractiveness of the
sharing platform, yet it is not a sufficient condition for actual resource sharing.

3.2. ENGINEERING INCENTIVES FOR SOCIAL CLOUDS

67

Only if the providers actually adhere to their offers the sharing makes sense. That
is, resource providers should have the incentive to fulfill their (voluntary) obligations and really provide the services. This is also true for resource consumers, as
they should be incentivized not to damage other users’ resources, e.g., not store
illegal of harmful data on other users’ machines or use offered VMs to execute malicious code.

Social Service Level Agreements There are several important aspects how a desired social behavior can be induced in platforms such as Social Clouds. One option is through the design of (resource or trading) agreements as well as their enforcement. Unlike the design of these components in other markets for electronic
resources or services where agreements are often described by Service Level Agreements (SLAs) and enforcement is achieved through penalties when SLAs are violated, other aspects have to be considered in a socially-oriented scenario such as a
Social Cloud. As the sharing of resources is, in its essence, voluntary, standard SLA
and penalty approaches are not suitable. This calls for the development of a new
type of SLA that considers these issues (see e.g. Michalk and Haas (2011)). One of
the premises of a Social Cloud is that real-world relationships should be utilized in
the resource sharing mechanism, yet the sharing and trading of resources should
have no negative effect on these real-world relationships. Hence, the specification
of resources to be shared through hard SLAs may not be an ideal approach as (volunteer) consumption of resources is, in many cases, not business driven. Furthermore, enforcement through monetary or other penalties can be counterproductive.
For example, if a friend fails to deliver the agreed-upon resource this is not comparable to when a company fails to deliver business-critical services. Hence, the
design of SLAs that are able to address the mentioned specifics of sharing in social
contexts is a necessary task, which however is outside the scope of this work.

Trust and Reputation Systems Another attempt at fostering social behavior in
online (sharing) communities is the design of a trust or reputation system (see
Jøsang et al. (2007) for an overview of related models). By giving users the ability to
provide feedback about other users’ actions, the trust and reputation schemes aim
to discourage malicious behavior through indirect enforcement. In other words,
even though users might not have interacted directly beforehand, the trust and
reputation system allows them to obtain information about each others’ previous

68

Incentives in Social Clouds

behavior. Examples for such trust models are SocialReGreT (Sabater and Sierra,
2002), EigenTrust (Kamvar et al., 2003), PeerTrust (Xiong and Liu, 2004), and PowerTrust (Zhou and Hwang, 2007), all of which are focused on establishing a reliable
trust rating for (potentially anonymous) user-based networks. Another example is
Petri et al. (2012) who apply a feedback-based trust model to P2P cloud communities, a setting very similar to a Social Cloud.
From the viewpoint of the incentive scheme, the second and third stage are closely
intertwined. Incentive schemes usually have the goal to increase participation or
encourage certain actions in a system, thereby inherently focusing on the second
stage. However, as the incentive scheme prescribes rules in the system (such as
which interactions are possible based on the contribution score of a user), these
rules might affect the behavior of users in the third stage as they determine the
potential action of a user. From a game-theoretic viewpoint on incentive systems,
besides the specification of agreement and enforcement structures, the concept of
strategically acting users has to be taken into account as well. In this context, users
are said to be acting strategically if they pursue some personal goal and take into
account the potential actions of other users while trying to achieve this goal. In
other words, users try to find the strategy that is most beneficial for themselves,
potentially irrespective of the effects of this strategy on other users or the system.

3.2.2. Factors Influencing the Participation Incentives
The importance of certain incentives throughout the user life-cycle depends on the
actual characteristic of the specific scenario. On one hand, design characteristics of
the platform affect user incentives, namely the relationship types of users given the
underlying social network, the applied market mechanism, and the specific application scenario. There are complex dependencies between these factors, which are
addressed in this section. On the other hand, users might have personal motivation to participate in a Social Cloud independent of its actual implementation. This
issue will be considered in the next section.

Application Scenarios and Market Mechanisms As discussed in Chapter 2.1,
there is a wide range of application scenarios for SCs, which are differentiated by
the specific types of resources that are shared (e.g., storage or VMs), the use case
(sharing among friends vs. sharing for a research project), and the organizational

3.2. ENGINEERING INCENTIVES FOR SOCIAL CLOUDS

69

setting of the scenario with respect to the participating users (private users vs. sharing between companies).
Considering these examples it is clear that the choice of proper incentives depends
on the actual application scenario, as users’ motivation might be quite different in
certain domains. In purely social or volunteer scenarios, currently implemented
systems such as myExperiment (De Roure et al., 2009) suggest that intrinsic motivations, fostered by Trophy or Reciprocation mechanisms, provide sufficient incentives to share resources. In contrast, extrinsic motivations (e.g., credits such
as nanos on nanoHUB.org) might be necessary to incentivize strategically acting
users in enterprise scenarios. A universal incentive system, thus, is not promising,
and the application scenario and its intricacies have to be taken into account in the
design of the system.

Relationship Types Besides the application scenario, other factors influence the
choice of an appropriate incentive mechanism. Although Social Clouds build upon
the inherent trust relationships between users, the level of trust depends on the
type and strength of users’ relationships. Ties between users can be different based
on the different social groups (family, friends, acquaintances, etc.), or the frequency
of interaction. The classification between strong and weak ties is an example for
such different strengths of relationships (Granovetter, 1973).

Colleagues, Acquaintances

Family, Close Friends

Others

Reciprocity-based
Mechanisms

Reciprocity-based, Trophy
or Credit-based
Mechanisms

Credit-based Mechanisms

Figure 3.2.: Example of Relationship Types and Resource Sharing Incentives

70

Incentives in Social Clouds

Consider, for example, the situation shown in Figure 3.2 where a user has different
types of relationships with other users of the underlying social network. Suppose
that three groups can be identified: Family and close friends, colleagues and acquaintances, and the remaining group of other people which include people where
the friendship only exists in digital form. The definition of such groups of friends
is already available, for example through relationship lists in Facebook or circles in
Google+. As one can imagine, a user might have different incentives to share with
users in each of these groups. Intuitively, one would rather share resources with
family members than with strangers without direct personal connections. Consider, for example, computation or storage of sensitive data. A user most likely
will trust family members or close friends to be reasonably careful with data (given
certain technical prerequisites), but this may not be the case with purely online
“friends”.
Therefore, it is useful to distinguish the incentives to share and the expected rewards with respect to these groups. Whereas users might share altruistically with
users from the first group and do not expect direct compensation due to the high
level of trust, the weak relationships in the third group might require direct rewards, such as receiving virtual credits, for the sharing of resources between these
groups. In practice, however, it is a nontrivial task to infer the relationship type
or strength of users based on interaction data within a social network (Xiang et al.,
2010). For example, close friends might not use social network platforms to interact frequently as they have other means of communication. Hence, the incentive
design of a Social Cloud should offer the users flexibility in the choice of how resources are exchanged and if some form of compensation is required.

Market Mechanisms Another factor that influences user incentives is the choice
of market mechanisms, which can also both depend on the application scenario and
on the relationship type (and thus the level of trust) between the users. Consider,
again, Figure 3.2. Building on the previous reasoning, certain market mechanisms
seem to be more appropriate for trading with different types of relationships. For
resource sharing with users of the first group, purely volunteer or reciprocal mechanisms may be most appropriate as the closer relationships might infer a sense of
reciprocation. In this setting a high level of trust is present and users expect no immediate reward in return or trust other users that they will share resources in the
future. On the other hand, for exchanges with users of the second group Trophy,
Reputation or Credit mechanisms may be adequate and offer certain advantages.

3.2. ENGINEERING INCENTIVES FOR SOCIAL CLOUDS

71

For example, reputation can compensate the lower levels of trust between the users
by providing additional information about other users, and credit systems would
allow users to accumulate credits which they can use for resource consumption.
Finally, due to the lower trust levels between the user and members of the third
group, credit-based systems or other monetary mechanisms might provide the only
form of reasonable incentives to engage in trading (see e.g. the multitude of creditor activity-based incentive schemes in anonymous P2P sharing communities summarized in Section 3.1).

3.2.3. Design Implications for Social Clouds
This section points out several important aspects concerning participation incentives in Social Clouds. First, it is necessary to obtain an understanding of the relevant motivations of potential participants throughout the identified stages of the
participation lifecycle. Only with knowledge about different user types and their
respective motivations it is possible to design incentive schemes that consider the
important, platform-specific intricacies. Second, the relevance of certain incentives
is affected by several factors such as the application scenario or the exchange mechanism. The interplay between the different user types as well as their interaction
with the incentive scheme and the exchange mechanisms implemented on the platform can be quite complex and hard to predict. This should be acknowledged in
the design of incentive schemes, and a design process utilizing sensitivity analyses
with respect to user type distributions can provide useful additional insights.
In order to address these aspects and challenges, the next section aims to get an
understanding of the relevant user motivations and incentives to participate in a
Social Cloud. In addition, Section 4.2 of the next chapter shows how the simulation tool presented in the last chapter is a complimentary methodology in the
design of Social Clouds. By simulating the introduction of an incentive scheme,
the non-trivial effects of that scheme on the different user types of the system will
be studied.

72

Incentives in Social Clouds

3.3. Identifying Relevant User Incentives
The previous sections introduced the various incentive problems that a market designer for a resource sharing platform needs to consider. Whereas these considerations were mainly driven by theoretical work and empirical observations of similar
systems, it is necessary to identify the relevant incentives for users of a particular
system and sharing scenario. For this reason, a small-scale web-based survey was
conducted as a first step to get a better understanding of incentives for potential
users. This survey can be seen as a pre-test of a more in-depth study of incentives
relevant for individuals or certain user groups.2 The focus of this section, therefore, is to describe the web survey and its results, aiming at identifying relevant
incentives for sharing platforms such as a Social Cloud.

3.3.1. Goals and Design of Web Survey
Identifying relevant incentives is necessary for all three participation stages as discussed earlier in Section 3.2.1. The survey discussed in this section focuses on the
first two stages, as the last stage (incentives for social sharing behavior) is inherently dynamic and requires feedback from participants that have used the system
in practice. Being closely related to research question 1.1, the goal of the survey is
to answer following questions:
1. What types of resources have users shared previously, with whom did they
share them, and what are the relevant motivations for previous sharing?
2. What are relevant incentives to join a sharing platform and actively share
resources? How does this depend on the previous experience, the sharing
scenario (private vs. professional networks), the use case (sharing of storage)
as well as the personality type?
The first goal was to identify the previous experience of the survey participant with
other sharing systems. This includes questions about the resources that have been
shared, the platforms that have been used as well as the user groups the resources
have been shared with. Such information is important for a system designer as the
2 Ideally,

such a survey would be coupled with a prototype of the system in question, and target
the users of the prototype. As the technical SC prototype was not ready to deploy at the time of
the survey, this was unfortunately not suitable.

3.3. IDENTIFYING RELEVANT USER INCENTIVES

73

(subsequently asked) sharing incentives might be different for users based on their
previous experience. The second goal was for users to provide information about
their perceived importance of certain incentives, such as monetary compensation
or altruism.
Together with the additional data gathered during the survey, which besides the
mentioned previous experience also includes demographic data and a short personality test, this allows for an exploratory analysis of the importance of relevant
incentives for certain user types. To limit the survey length, the Ten Item Personality Inventory (TIPI, Gosling et al. (2003)) was used instead of the longer standardized questionnaires to assess personality types. Additionally, two sharing scenarios
were distinguished, sharing in a private network vs. sharing within a professional
network, in order to identify potential differences based on the sharing scenario. In
a private network, sharing occurs between friends, whereas the focus of the professional network is to share between colleagues, co-workers and classmates.

Survey Implementation The survey was implemented in English and German
using SurveyMonkey3 and consisted of three main blocks. The first block considered questions about previous sharing behavior. If the survey participant previously shared resources in similar settings, questions about the previously used
platforms, shared resources, and the motivations to share these resources are asked.
Given that users, independent of their previous sharing experience, are interested
in sharing resources in similar settings, the second block studies the motivations
to participate and actively provide resources in two hypothetical sharing scenarios. The goal of this block is to identify if there are differences in the importance of
certain motivations between the scenarios, which considered sharing in a private
(friend) network compared to a professional (business) network. Finally, the last
block concluded the survey with the aforementioned personality and demographic
questions.
Given that participants of the survey do not necessarily have experience with previous resource sharing, or are not even interested in sharing resources, the survey
included two additional questions that acted as filters in the processing logic. The
first filter asked participants if they previously have shared resources, and either
directed them to the first block of questions, or to the second filter question. This
second filter asked survey participants if they are in general interested in sharing
3 http://www.surveymonkey.com

– last accessed May 2014

74

Incentives in Social Clouds

resources. With a positive response, the participant was directed to the second
block of questions, otherwise to the demographic questions. Appendix A shows
the logical structure of the survey (Figure A.1), as well as the survey questions.
The survey included closed questions with a given set of answers to choose from
as well as open questions. For the questions involving assessment or evaluation, a
6-point Likert-scale was used.4

3.3.2. Evaluation
Having introduced the survey goals and logic, this section presents the results of
the online survey. In particular, the results considering the research questions are
addressed. Especially, the following demographic statistics show that the participants reflect the local (student) community to a high degree. Hence, it is not a
representative sample of the population, which is why the results and potential
implications of the survey have to be interpreted carefully.

General Statistics and Demographics The survey was conducted over a period
of two months from December 2012 to February 2013. In total, 172 people participated the survey, whereof 126 fully completed it (73.3%). Of the 126 completed
responses, 5 are not considered for the evaluation due to not meeting the data validity criteria. In particular, these five responses contained participants that could
not have possibly read all answered questions thoroughly in time (lower time limit
1 minute if all blocks were answered), checking the same item on the Likert-scale
for all answers, or being extremely contradictory in their TIPI answers (putting
themselves on the opposite extremes for the personality factors). This leaves 121
fully completed and valid responses which are considered for the subsequent evaluation.5

4 There

is currently no consensus if a scale with even or odd options should be preferred. On one
hand, using a scale with an even number of options cannot represent a truly neutral opinion, and
by forcing the participant to choose a non-neutral option might thereby bias the results (Schnell
et al., 2011; Garland, 1991). On the other hand, having an odd number of options and using the
midpoint can be used to avoid a tendency or to get through the survey as quickly as possible,
potentially without reading or thinking about the questions. In such a case it could not represent
a truthful statement and hence bias the result as well (Weijters et al., 2010).
5 Including the not considered 5 responses does not change the presented results significantly.

3.3. IDENTIFYING RELEVANT USER INCENTIVES

75

Considering language, 86.0% of the participants completed the German version of
the survey, and 14.0% the English version. This indicates that the main pool of
participants was probably sourced locally. Nearly three quarters of respondents
were male (73.9%), which also indicates that mostly local people participated as
this reflects the local gender distribution at the Karlsruhe Institute of Technology.
The majority of participants were in the age groups of either 20-25 (46.2%) or 2635 (35.3%). Students make up the majority of participants (61.4%), followed by
employees (30.3%). 8.3% either list another profession or preferred to not answer
this question. In addition, over half of the respondents have a university degree
(55.4%), and 36.2% have a high school diploma or “Abitur”.

Previous Resource Sharing Experience The first block of the survey asked participants about their previous experience in sharing resources online. This block was
answered by 119 out of the 121 participants, meaning that 98.3% of participants
had such previous experience.
8VHU*URXSV3UHYLRXV6KDULQJ
)ULHQGVUHDOOLIH



)DPLO\



&ODVVPDWHV



&ROOHDJXHV



)ULHQGVRQOLQH



)ULHQGVRI)ULHQGV
2WKHU
















3HUFHQWDJH

Figure 3.3.: User Groups with which Resources were Previously Shared Online

First, the participants were asked which online platforms they previously used to
share resources. For the given sample, the majority of participants used email,
DropBox or other (private) social networks to share resources online (see Figure A.2
in Appendix A). Other Cloud- or P2P-based platforms are also frequently used, for
example Google Drive, P2P platforms and own servers. Interestingly, professional
networks are used much less frequently than private social networks to share resources online (13% vs 74%). Besides the platform for sharing, participants were
asked which type of resources they previously shared (Figure A.3 in Appendix A).
Whereas resource types such as files, pictures, and music are common as expected,

76

Incentives in Social Clouds

it also shows that both lecture notes and sample solutions are shared frequently,
which can be explained by the high number of students in the given sample. Furthermore, storage is shared by approximately 24% of the participants, which is
relevant for another question later in the survey.
5HOHYDQFHRI,QFHQWLYHVIRU3UHYLRXVO\6KDUHG5HVRXUFHV


5HOHYDQFH






'LUHFW5HTXHVW

$OWUXLVP
+HOSIXOQHVV

2ZQ%HQHILW

)LQDQFLDO
2WKHU
5HSXWDWLRQ
&RPSHQVDWLRQ&RPSHQVDWLRQ

Figure 3.4.: Incentives for Previous Online Resource Sharing

Figure 3.3 shows the user groups with which the participants have shared resources
online. Sharing among direct friends, relatives, and classmates is very common,
whereas fewer participants share resources with users they only know indirectly
or online. This can be an indication that the necessary level of trust to facilitate
sharing is higher for these user groups.
Finally, the last set of questions in the first block considered the incentives that were
important for participants in their previous online resource sharing experience.
The participants were asked to rank the relevance of certain incentives, where 1 on
the Likert-scale corresponds to “not important” and 6 is “very important”. The results are shown as boxplots in Figure 3.4. The data reveals that direct requests from
other users, general altruism and helpfulness, as well as the own benefit of sharing are considered the most important incentives by the majority of users. Social
reputation and other forms of compensation are less important for most users, and
financial compensation is considered least important. These results indicate that
direct compensation is not necessarily a driver in resource sharing, participants
rather seem to include strategic considerations such as reciprocity in their decision
to share resources online, as shown in the high importance of “own benefit”.

3.3. IDENTIFYING RELEVANT USER INCENTIVES

77

Incentives for Registration and Sharing After considering the experiences of
previous online resource sharing, the second block of questions in the survey studied the importance of certain incentives in hypothetical resource sharing scenarios.
This covered the first two stages of the sharing lifecycle, namely incentives to register as well as incentives to actively participate. Furthermore, the scenarios distinguished between sharing in private social networks (such as Facebook) and professional social networks (primarily used for business and professional networking,
such as LinkedIn), to investigate if users have different incentives to participate in
resource sharing based on the type of relationships they have with other people in
the network.
1HWZRUN5HJLVWUDWLRQ
1HWZRUN



3ULYDWH
3URIHVVLRQDO

6FRUH






5HTXHVW&ORVH 0RQHWDU\
)ULHQGV
&RPSHQVDWLRQ

5HSXWDWLRQ

&XULRVLW\)XQ

+HOSIXOQHVV

2ZQ%HQHILW

(a) Registration Incentives
1HWZRUN3DUWLFLSDWLRQ
1HWZRUN



3ULYDWH
3URIHVVLRQDO

6FRUH










5HTXHVW&ORVH 0RQHWDU\
)ULHQGV &RPSHQVDWLRQ

5HSXWDWLRQ

&XULRVLW\)XQ +HOSIXOQHVV

2ZQ%HQHILW

(b) Participation Incentives
Figure 3.5.: Survey Results for Registration and Participation Incentives in Private and Professional Networks

78

Incentives in Social Clouds

There are several interesting results that can be seen in Figure 3.5. Considering sharing resources with other users in a private network (e.g., non-colleague
friends), the relevant incentives for registration and active participation on the platform are quite similar, as shown in Figures 3.5a and 3.5b. The most important incentives are requests from close friends, as well as the own benefit from sharing
(e.g., expected reciprocity). Furthermore, curiosity and helpfulness also are important for many participants, yet less important than the previously mentioned
incentives. Financial compensation is, similar to the case of previous resource sharing in Figure 3.4, the least important incentive for most participants.
The results were also compared using a Wilcoxon signed-rank test as the data was
not normally distributed. The statistical evaluation showed that in private networks, monetary compensation (p = 0.01), helpfulness (p = 0.01), as well as curiosity/fun (p = 0.05) are considered more important for the participation stage.
In professional networks, a statistically significant difference in the importance between registration and participation incentives could not be detected.

The Difference of Private and Professional Networks Users might perceive private networks differently to professional networks. Hence, it is interesting to see
if the relevance of the considered incentive categories are different for users across
these two types of networks. The comparison of incentives in private and professional networks in Figures 3.5a and 3.5b reveals interesting results. In both network
types, the personal benefit (expected reciprocity) is the most important incentive
for both registration and active participation, even more important than direct requests. Furthermore, curiosity and fun plays a lesser role in case of professional
networks, and financial compensation is slightly more important in this context
(which is in line with the previous arguments in Section 3.2.2).
As discussed previously, the relative ranking of the incentive categories between
the network types is similar, yet there also seem to be differences in the level of
importance. Comparing the differences of participating in private networks compared to public networks, a Wilcoxon signed-rank test is used to compare the
differences between the two types of networks as the data was not normally distributed. For the different types of incentives to join such a platform, the test detected significant differences in Monetary Compensation (more relevant in professional networks, p < 0.001), Reputation (more relevant in professional networks,
p < 0.001), and Curiosity/Fun (less relevant in professional networks, p < 0.001).

3.3. IDENTIFYING RELEVANT USER INCENTIVES

79

Considering the incentives to actively provide resources in private versus professional networks, the Wilcoxon signed-rank test found significant differences in Direct Request (less relevant in professional networks, p = 0.001), Monetary Compensation (more relevant in professional networks, p = 0.012), Reputation (more
relevant in professional networks, p < 0.001), Curiosity/Fun (less relevant in professional networks, p < 0.001), and Helpfulness (less relevant in professional networks, p = 0.029).
As the sample size is rather small, these results are merely a tendency, yet confirm some of the expectations and reveal several interesting facts. The results show
that in general, ones own expected benefit as well as direct requests are the most
important incentives for both private and professional resource sharing networks.
However, there seem to be subtle differences originating from the different context
of sharing, where sharing with friends tends to place a higher emphasis on curiosity and helpfulness, and sharing with colleagues leads to a slightly higher emphasis
on monetary compensation.

Sharing of Storage Space The last set of questions in the second block considered
a scenario in which storage space is shared between users, allowing other users to
store data on the participant’s machines. The questions were framed in a way that
indicated that such storage would be completely safe for the storage providers due
to security and encrypting techniques. As storage sharing is a relevant use case
and already implemented on certain platforms, the aim of this question set was to
identify the necessary user groups and relationship types for which participants
might be willing to share storage resources.
Considering the user groups that participants are willing to share storage with, the
results show that besides a small minority that either is not willing to share storage
at all or willing to share with everybody, most participants would be willing to
share with family, relatives and close friends (see Figure A.4 in Appendix A). About
30% are also willing to share with classmates and colleagues, yet only a small subset
of participants would be willing to share storage with friends of their friends. This
also indicates that most participants require a certain level of trust towards the
other user in order to allow sharing of storage space. In particular, assumptions
about trust transitivity have to be handled carefully as the results for friend-offriend relationships indicate that the perceived level of trust considerably decreases

80

Incentives in Social Clouds

for indirect connections. This is confirmed in Figure A.5, which indicates that users
specifically require close relationships for sharing of electronic storage space.

The Influence of Personality Type on Sharing Incentives After the previously
discussed questions about incentives for online resource sharing, the survey closed
with some short demographic questions as well as the TIPI. The aim of including
the TIPI questions was to potentially identify if certain incentives are more relevant
for certain personality traits. Figure 3.6 presents an overview of the participants’
personality profiles based on the ten TIPI questions. The correlation tables with the
extended results can be found in Appendix A.



6FRUH






([WUDYHUVLRQ

$JUHHDEOHQHVV

&RQVFLHQWLRXVQHVV (PRWLRQDO6WDELOLW\

2SHQQHVV

Figure 3.6.: Survey Results for Ten Item Personality Inventory (TIPI)
Table 3.1.: Comparison of Survey TIPI Scores with original TIPI scores (Gosling et al., 2003)
and TIPI-G (Muck et al., 2007)
Data

Sample
Size

Extrav.

Agreeabl.

Consc.

Emotional
Stab.

Openness

Survey
TIPI
TIPI-G

121
1126
175

4.30 (1.38)
4.56 (1.48)
4.87 (1.21)

4.72 (1.05)
5.26 (1.12)
5.20 (0.95)

5.23 (1.16)
5.47 (1.13)
5.85 (0.93)

5.01 (1.14)
4.85 (1.45)
5.10 (1.20)

5.17 (1.03)
5.43 (1.06)
5.49 (0.97)

As the sample size is rather small, and the TIPI cannot provide as thorough a personality trait classification as the longer Big Five Inventory can (John et al., 1991),
the results of the correlation analysis have to be interpreted with care. Table 3.1
compares the mean and standard deviation of the answers to two other data sets,
the original TIPI data set (Gosling et al., 2003) as well as a German adaption of the
TIPI questionnaire (TIPI-G, Muck et al. (2007)). Although there are small differences, overall the values are similar to the other data sets. Unfortunately, even if

3.3. IDENTIFYING RELEVANT USER INCENTIVES

81

there are tendencies in the data, the sample size is also too small to conclusively
identify significant correlations between personality traits and the importance of
certain incentives except for some cases.
Despite these issues, there are some results that are noteworthy to discuss. First of
all, there is an expected significant positive correlation between ’Extraversion’ and
’Openness’. Further positive correlations are between ’Extraversion’ and ’Conscientiousness’ as well as ’Emotional Stability’, and between ’Conscientiousness’ and
’Emotional Stability’ (see e.g. Tables A.1 and A.2).
Considering TIPI and relevant previous sharing incentives, there is a significant
positive correlation between ’Conscientiousness’ and direct requests as well as general helpfulness (see Tables A.1 and A.2). For sharing in private networks, both in
case of registration and participation incentives ’Extraversion’ is significantly positively correlated with monetary compensation, reputation, own benefit and curiosity/fun, yet only in the participation scenario there is a significant correlation
between ’Extraversion’ and requests of close friends. Additionally, ’Agreeableness’
is positively correlated with helpfulness (see Table A.3). In professional networks,
also both for registration and participation incentives, ’Extraversion’ is significantly
positively correlated with requests from friends, own benefit, curiosity/fun, and
reputation. Furthermore, in these cases ’Openness’ is also significantly positively
correlated with curiosity/fun and helpfulness, indicating that these incentives are
more relevant for people considering themselves more open (see Table A.4).
In summary, these results provide a good basis for further investigation of relevant incentives, especially targeted at more specific scenarios. The insights gained
by this short survey are helpful in getting an overview of participants’ attitudes
towards sharing, their potential willingness to share with indirectly connected network users, and different relative importance of incentives based on the type of
relationships between the users.

3.3.3. Discussion
The survey results shed light on the relevance of certain incentives for sharing in
a Social Cloud. In particular, for the registration and participation on such a platform, non-monetary incentives (such as helpfulness, curiosity, fun, and reputation)
have a higher relevance than monetary compensation, which is similar to findings

82

Incentives in Social Clouds

from literature. Considering relationship types, the sharing of resources with close
friends is more prevalent than the sharing with acquaintances or indirect friends.
Table 3.2.: Relevance of Incentives for Private and Professional Networks, Mean and Standard Deviation
Incentive
Registration
Request Close Friends
Monetary Compensation
Reputation
Curiosity/Fun
Helpfulness
Own Benefit
Participation
Request Close Friends
Monetary Compensation
Reputation
Curiosity/Fun
Helpfulness
Own Benefit

Stat. Difference
(p=0.05)

Private

Professional

4.69 (1.26)
2.34 (1.39)

4.45 (1.27)
2.88 (1.64)

2.66 (1.40)
4.13 (1.28)
3.99 (1.25)
4.76 (1.15)

3.35 (1.55)
3.43 (1.44)
3.98 (1.29)
4.81 (1.24)




4.85 (1.24)
2.67 (1.61)

4.50 (1.29)
3.04 (1.73)




2.83 (1.49)
3.92 (1.37)
4.31 (1.06)
4.74 (1.17)

3.31 (1.52)
3.46 (1.46)
4.07 (1.20)
4.70 (1.25)







The results also show differences in the relative importance of certain incentives
with respect to the sharing scenario. Table 3.2 provides an overview of the difference in said relevance. Monetary compensation overall is of lesser relevance than
other incentives, yet more important in professional settings. In contrast, curiosity/fun is of higher importance in private settings.
The identified relevant incentives also have to be considered in the design of a Social Cloud. For example, the relevance of direct requests implies that corresponding
interaction capabilities between users should facilitate such direct requests. Considering resource sharing in social contexts, the findings show that the specific
sharing scenario has to be taken into account for the incentivization of users. In
particular, the importance of non-monetary incentives emphasizes the necessity to
implement allocation mechanisms that are not based on monetary transactions or
compensations.

3.4. SUMMARY

83

3.4. Summary
This chapter identified and discussed the explicit need to address user participation incentives in Social Clouds. Its aim was to address research question 1.1. Section 3.2 proposed a model in which three participation stages are distinguished,
namely User Discovery and Registration, User Participation, and Social Behavior
in Resource Sharing. Additionally, factors that influence the importance of incentives, such as the application scenario, relationship types and the applied market
mechanisms, were identified. An online survey was conducted to identify potential
differences in the importance of incentives within different stages, and its results
discussed in Section 3.3. The results show that non-monetary incentives such as
altruism, fun, or expected reciprocity are more important then monetary compensation, which is in line with the Social Cloud assumptions. Regarding factors that
influence the importance of incentives, the results indicate that for the given scenario and participation pool, the setting of the network (private vs. professional)
influences the importance of certain incentives such as monetary compensation.
Furthermore, users prefer to share resources along stronger relationship types, i.e.,
favor sharing with friends over sharing with colleagues or friends of friends. In
contrast, for the given scenario the importance of the considered incentives was
similar in both the registration and participation stage.
As discussed before, the identification of user participation stages as well as the
relevant incentives in these stages are a necessary first step in the design of an incentive scheme. As a next step, this knowledge can then be applied in the design
of an incentive scheme tailored to the given scenario. Before adapting and implementing the incentive scheme on the platform itself, a simulation-based approach
can be leveraged to study the sensitivity of said scheme on changes in the system,
e.g., a change in user type distributions. Therefore, the next chapter presents two
case studies that show how incentive and contribution schemes can be evaluated
with respect to different user characteristics.

Chapter 4.
Designing Incentive Schemes and
Co-operative Infrastructures
“It is well known in the public goods literature that in the absence of outside incentives the
individually rational allocation of resources in such an environment will, in general, be
less than the socially optimal outcome.”
(Krishnan et al., 2006)

T

HE identification of participation stages and the relevant incentives therein is
the necessary first step in the development of an incentive scheme tailored to
the specific needs of a Social Cloud. Based on these insights, incentive schemes can
be developed that incorporate the findings. In practice, the impact of such incentive
schemes needs to be studied to assure that the goals of the incentive scheme are met
and that the incentives are set correctly.
As mentioned before, there are several ways how incentive schemes can be designed and evaluated, such as implementing the proposed incentive scheme and
test it through a prototype system with a test user group, or deriving predictions
based on analytical models. Complementary to these approaches, a simulationbased study can be particularly helpful when evaluating through a prototype is
not possible or feasible, e.g., in the early development stages of a system or when
changes to the system are expensive and the effects need to be predicted beforehand. An example for such an application is the prediction of the effects of changing user groups and user behavior, and their impact on the efficiency of the incentive scheme. Simulation-based approaches have been advocated before for use in

85

86

Designing Incentive Schemes and Co-operative Infrastructures

the design process of incentive systems in online communities, see e.g. Mao et al.
(2007) and Ren and Kraut (2011). Hence, such a simulation-based approach in the
engineering of incentive schemes is the focus of this chapter.
This chapter illustrates how incentive schemes can be designed and studied using a simulation-based approach by presenting two case studies. Specifically, this
chapter aims to answer research questions 1.2 and 1.3. Section 4.1 describes the
user model that is used in both case studies. In Section 4.2, the first case study develops an incentive scheme with the aim to encourage resource contribution from
users. During the case study, the (dynamic) effects of the incentive scheme on the
overall system, as well as on different user types, are evaluated. The second case
study, which is presented in Section 4.3, models a platform for which the necessary
resources are co-operatively provided by the platform users themselves. Specifically, it aims to study the suitability of different contribution schemes (i.e., how
users contribute to the co-operative infrastructure) for the considered platform.
The chapter concludes with a discussion of the case studies.

4.1. User Model
Users in the case studies are described by several characteristics, which include:
• Resources: Users have a resource endowment, which can be of several resource types. They can offer parts of their resources and may request resources from other users. The resources are generally described by amount
and type.
• Objective: The goal of the users, in general, is assumed to be the maximization
of their utility function. In this case, the objective is generally to maximize the
(positive) difference of benefits (e.g., through consumption of resources) and
costs (e.g., for offering resources).
• Availability: As users are probably not available all the time, their availability
can be described by respective distributions. For example, Javadi et al. (2011)
identify several user groups for the Volunteer Computing project SETI@home
with distinctive distributions for availability and unavailability intervals.
For ease of reference, Table 4.1 provides an overview of the parameters that are
used in the case studies.

4.1. USER MODEL

87

Table 4.1.: Model Parameters
Parameter

Description

β
λ
p
ωi,r
r
ρi,r
si,t
σi,r

Convexity parameter
Degree of altruism
Relative price of providing or sharing resources
Resource endowment of user i and resource type r
Resource type
Percentage of resources of type r that user i shares
Contribution score of user i at time t
Percentage of resources of type r that user i reserves for own
purposes
Minimum amount of resources that user i reserves for own resources
Amount of resources that user i reserves for own purposes
Amount of resources that user i shares
Utility parameter from consuming resources from other users

σi,min
Πi,o
Πi,s
τi,r

Resources In the case study model, each user i has an endowment of (computing)
resources ωi,r available, where r ∈ R indicates the resource type. They can use them
either for (self-) consumption or provision them to other users or the infrastructure.
The percentage of resource type r that the user consumes or reserves for their own
purposes is denoted by σi,r (this includes not using the resources at all, i.e., idle
resources). Similarly, δi,r is the percentage of resource type r that user i shares with
other users (or the infrastructure in the second case study).
The cost for resource provisioning is indirectly determined through the parameter
p: the higher p, the higher the relative cost of providing resources compared to
consuming them or leaving them idle. As users only have a limited amount of
resources available, for each user i the following endowment constraint has to be
fulfilled:
(4.1)
σi,r + pρi,r = 1
This indirect approach of modeling costs was chosen for following reasons: first,
modeling an explicit cost function requires assumptions about its components and
their relation. For example, although there are studies that model energy costs for
computer systems, their exact form (especially for different resource types) is hard
to determine (see e.g. Elnozahy et al. (2003)). In addition, the detailed measuring of energy consumption and costs requires additional equipment such as smart
meters, which might not be available to all users. Second, one might question if
explicit cost considerations have a major influence on the decisions of users in social contexts. While this might be true for anonymous systems such as P2P, it is not

88

Designing Incentive Schemes and Co-operative Infrastructures

clear if users explicitly take into account the costs for supporting a resource request
of a (potentially indirect) friend.

Objective Using the standard economic approach, users are assumed to optimize
a certain utility function. In contrast to classic economic theory, however, the utility
function is not modeled as being purely dependent on the consumption and costs
of resources. As the concept of social or other-regarding preferences and a multitude of
studies have shown, for their decision making users might take other users’ behavior and consumption into account (see e.g. Fehr and Schmidt (2000); Bolton and
Ockenfels (2000); Andreoni and Miller (2002); Falk and Fischbacher (2006); Fehr
and Schmidt (2006)). Hence, the case studies model a user’s decision through a
utility function that incorporates these aspects.
Specifically, the user model makes use of utility functions studied by Andreoni and
Miller (2002). They provide general utility functions that can explain findings from
economic experiments in which a significant amount of subjects exhibit altruistic
behavior, which is in contrast to the traditional economic principle of self-interest
and rational utility maximization.1 More specifically, altruistic behavior in this case
means that a user “is willing to sacrifice own resources in order to improve the well
being of others” (Fehr and Schmidt, 2006, p.620). This is the case when users provide resources for others or the platform infrastructure, as they incur costs (not
being able to use the resources themselves) but their contribution increases the performance of the system. Andreoni and Miller (2002) show that through correct
parametrization of the following utility function (Equation 4.2), altruism can be included in economic decision making. Their approach is used for two reasons: 1)
the given setting closely resembles a public goods game where users incur costs
for providing resources to the public (other users or the co-operative infrastructure), and in this case self-interest and altruism are often used as the most relevant
parameters that characterize the behavior of users (see e.g. Fehr and Fischbacher
(2003)). 2) Their utility function is able to capture various forms of utility functions,

1 Based

on certain assumptions, it is common in economics to describe the preferences that users
have with respect to an outcome (such as resource allocation) with a utility function that implements these preferences (Andreoni and Miller, 2002). Andreoni and Miller (2002) also showed
that this utility function is able to explain empirical data of several laboratory games which are
conceptually very similar to the considered scenario, specifically that a user’s actions affect the
utility of other users.

4.1. USER MODEL

89

from classic substitutive and Leontief2 to convex utility functions, making it very
flexible.

1/β
Ui (σi , ρi ) = ∑ Πi,s (σi,r , τi,r ) β + λΠi,o (ρi,r ) β

(4.2)

r

Here, Πi,s denotes the self-consumption of resources of user i, and Πi,o denotes the
provisioning of resources. As mentioned before, σi,r is the percentage of resource
type r that the user reserves for own purposes, and ρi,r the percentage of resource
type r that they share with other users. β is a parameter that defines the convexity
of the preferences, and λ defines the degree of altruism.
Besides consuming their own resources and giving their resources to other users,
users gain a certain utility from successfully served resource requests, denoted by
τi,r . For simplicity, this is assumed to be linear in the percentage which is served by
other users, with limits [0, 1]. In other words, the more of their resource requests
are fulfilled by other users, the higher the respective utility. τi,r is included in the
utility function as shown in Equation (4.2) because, in some sense, the utility from
consuming other resources can be considered similar to consuming one’s own resources. Furthermore, note that τi,r does not directly depend on (and hence cannot
be directly influenced by) the decisions of user i, as it depends on the number of
resources that the other users provide.
Furthermore, each user belongs to a certain utility type as identified by Andreoni
and Miller (2002). The different types and the parameters of the utility function
are summarized in Table 4.2. Besides three types with different values for β and λ
(types 4-6), types 1-3 describe special shapes of the utility function. Type 1 models selfish users who only receive utility from resource consumption, having no
inherent incentive to provide resources. Type 2 models a classic Leontief utility
function, and type 3 models the case when self-consumption and provisioning are
perfect substitutes.
Using the individual contribution to the system as the altruistic parameter is not
the only possible way of modeling “social” behavior; alternatives can be studied
as well. For example, instead of individual contributions, the average utility of
other users could be used. Other possibilities are to use the average contribution
2 In a Leontief utility function, the utility depends on the minimum value of the given components:

U ( x1 , x2 ) = min { x1 , x2 }. See e.g. (Mas-Colell et al., 1995, p. 49).

90

Designing Incentive Schemes and Co-operative Infrastructures

Table 4.2.: List of Potential Utility Function Types based on Andreoni and Miller (2002)
Type

Function

Size

λ

β

Description

1
2
3
4
5
6

Ui = ∑λ Πi,s,λ
Ui = ∑λ min {Πi,s,λ , Πi,o,λ }
Ui = ∑λ Πi,s,λ + Πi,o,λ
see equation (4.2)
see equation (4.2)
see equation (4.2)

0.227
0.142
0.062
0.245
0.162
0.162

0.319
0.529
0.736

0.621
-0.350
0.669

Selfish
Leontief
Perfect Substitutes
Weak Selfish
Weak Leontief
Weak Perfect Substitute

and/or consumption amount in the system, or the average amount of successfully
satisfied requests.3 While undoubtedly interesting, these aspects are left for future
case studies.
Based on their utility function, each user will choose their individual optimal values for ρi,r and σi,r depending on their level of altruism and the convexity of their
preferences. Note, however, that a user might wish to reserve a minimum level
of their endowment, σi,min , for other usage (such as private use or sharing with
friends). In the model, this would be equivalent to reserving a certain percentage
of ωi , and only have (1 − σi,min ) ωi potentially available for usage. As this is equivalent to an adjustment of user i’s available resource endowment, in order to avoid
making the model unnecessarily complex this aspect is omitted in the remaining
model.

4.2. Case Study: Designing Incentive Schemes
Having identified the participation stages of users as well as relevant incentives for
resource sharing in Chapter 3, the next step is the application of this knowledge
in the design of an incentive scheme for Social Clouds. To achieve this, the findings have to be transformed into a suitable incentive scheme, which is the focus of
this case study. It investigates the effects of an incentive scheme on the considered
sharing platform and its users to encourage the contribution of resources to fulfill on-demand computing requests of users. The case study is an example how a
simulation-based approach as described in Section 2.2 can be used in combination
3 Consider

the following simple example to explain why such an approach can make sense: a user
contributes a data set to a community or colleague and consequently a new discovery is made
which benefits that user or the community as a whole. In that case, the benefits for the other
users might influence the individual contribution of resources.

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

91

with the previously gained insights in the design and engineering of an incentive
scheme. The section is an extended version of Haas et al. (2012). It starts with an
outline of the scenario in Section 4.2.1, and describes the simulation model in Section 4.2.2. The case study is evaluated in Section 4.2.3 and the findings discussed
in Section 4.2.4.

4.2.1. Scenario
The case study builds on the work of John et al. (2011), and Thaufeeg et al. (2011)
who analyzed how the Social Cloud concept can be used to augment traditional
Volunteer Computing by leveraging social networks to encourage participation. It
builds upon the basic premises of volunteer computing that users contribute to
“worthy” projects, retaining the argument that if user communities leveraged their
social relationships as sources of resources, their accessible compute power would
exceed that of existing volunteer computing projects (John et al., 2011).4
In the given scenario, a user can request certain technical entities/resources that
are needed, e.g. for a specific scientific project. Examples could be a BOINC-like
project (Anderson, 2004) that mainly requires computing cycles as resources, or
more complex projects that require more than one resource type, e.g. databases,
memory, and so on. An exchange of virtual machines as presented in Chapter 6
is also possible. The request of the user is then advertised to other users in their
Social Cloud. These users decide whether or not to supply the required resources.
In general, the more links to other members a user has in the system, the higher the
potential supply of resources and thus the likelier it is that a request can be entirely
fulfilled. The main objective of this case study is to evaluate the effects of a trading
constraint (TC) on the system, specifically the effects on different user types.

4 This

is based upon the number of users in Facebook (1.19 billion monthly active users in 11.2013,
see http://tinyurl.com/fb-size), that an average user havs 190 friends see http://
tinyurl.com/fb-anatomy (last accessed May 2014), and the assumption that sharing also
occurs between friends of friends.

92

Designing Incentive Schemes and Co-operative Infrastructures

4.2.2. Modeling an Incentive Scheme
User Behavior
The user model builds on the utility function presented in Section 4.1. Without loss of generality, Πi,s and Πi,o can be general functions that describe how a
user’s utility depends on the consumption and provisioning of resources. This case
study, for simplicity, only considers Equation (4.2) with Πi,s (σi,r , τi,r ) = σi,r + τi,r
and Πi,o (ρi,r ) = ρi,r , i.e., the users’ utility is linearly dependent on the consumption
and provisioning percentage, respectively. Note that the components σi,r , ρi,r and
τi,r are normalized to the range [0, 1]. This is done to make different resource types
comparable in the utility function, assuming an equal importance of resource types
for the users.5
The optimal level of resource consumption and (through the endowment constraint) also for provisioning can be calculated through the derivative. Depending
on the relative price for resource provisioning, users select a certain percentage of
their resources that they contribute to other users. Assuming that the utility function can be segmented into a sum of functions (one for each resource type), and the
types are independent of each other, the optimal percentage can then be calculated
as:
∂Ui (σi , ρi )
=0
∂σi
−1

σi,r =

−β
β −1

λ 1− β p r
1+λ

−1
1− β

− τi,r
−β
β −1

∀r ∈ R

(4.3)

pr

Note that the maximization is only meaningful for either σi,r or ρi,r , as one determines the other, and τi,r only depends on the amount of resources that other users
provide and cannot be directly influenced by user i.6
While the percentage of provided resources as in Equation (4.3) might be individually optimal when utility maximization is considered to be the goal of the users,
5 The

model can be augmented to capture user-specific preferences for different resource types,
which is, for ease of understanding, omitted in this case study.
6 Note further that for utility type 5, restricting the range to [0, 1] would yield a theoretical optimum
of σ = 1, due to asymptotic conditions. However, as this is unrealistic, 1 is excluded from the
optimization, in which case equation 4.3 applies again.

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

93

one might ask if there are other, system-centric goals that are more in line with the
philosophy of the system. One such example could be the percentage of resource
requests that can be fulfilled.

Incentive Scheme
While a utility function only models the willingness to contribute, given assumptions or data on the distribution of user types and the respective parameters, a
sharing platform without additional rules might not induce the proper incentives
for each user group to participate in resource sharing. Hence, an additional trading
constraint is introduced as a type of incentive scheme in order to limit the effects of
free-riding and incentivize users to increase their resource contributions.
The trading constraint is similar to the incentive scheme described by Ranganathan
et al. (2004) and other contribution-based schemes. The main idea of the incentive
scheme is that each user receives a contribution score based on its contribution
to the Social Cloud, and each user can only trade with other users that have a
lesser or equal contribution score. In other words, if si (t) is the score of user i at
time t, a trade can only occur with users j for which s j (t) ≤ si (t). Increasing one’s
score allows the user to select from a broader base of other users to satisfy resource
requests, which affects the probability that resource requests are fulfilled (and thus
has an influence on the utility).
Ranganathan et al. (2004) argue that this is an effective incentive scheme in P2P
networks, leaving free riders with little options but to increase their participation
if they want to continue using the network. However, they did not study the effect
of the incentive scheme on the system itself, e.g., considering performance metrics
such as number of different files (resources) shared in the system. Different user
types, which might be adversely affected by the constraint, were also not considered.
For the evaluation this case study looks at two system performance measures:
1. The average utility of users
2. The number of satisfied requests
Intuitively, one might suspect that introducing a trading constraint reduces both
average utility (due to the utility of free riders being lowered) and successful allocations (as the supply for a request is shortened for users with lower contribution

94

Designing Incentive Schemes and Co-operative Infrastructures

scores). However, as users gain utility also from successful requests in the system,
increasing their contribution could increase the number of successful allocations
(more supply), and thus in turn increase their utility again. This case study evaluates which of these effects prevail.

Contribution Score The contribution score consists of two weighted components
as given in Equation (4.4).7 As a baseline, κ1 = κ2 = 0.5, is used, i.e., an equal
weighting between the contribution components. This can be changed in order to
shift the emphasis for different scenarios.
si ( t ) =

κ1 squantity,i (t) + κ2 sscarcity,i (t)
κ1 + κ2

(4.4)

The first component, squantity,i , is calculated based on the quantity of resources that a
user provides, giving an incentive to provide more resources to increase the contribution score. While this is, for itself, not necessarily beneficial for the system, as a
user can provide huge amounts of resources that are not needed by other users at
all, quality and relevance of resources can also be included. The quantity score of
a resource is determined through Equation (4.5). Over a certain time period (Tc ),
the maximum contribution of all users is determined, and the contribution of every
user is normalized by that maximum such that the range of the score is in [0, 1].

t
squantity,i (t) =

t− Tc

max
j

∑ ρi,r (τ ) dτ

r∈R

t

t− Tc

(4.5)

∑ ρ j,r (τ ) dτ

r∈R

The second component of the contribution score, sscarcity,i , is the scarcity of a resource,
which is determined through Equation (4.6) as the relative amount of requests vs.
offers. The higher the ratio, the scarcer a resource is, and providing scarcer resources yields higher contribution scores (with |Oi | being the number of resources
the user offers).
∑ ψr (t)
sscarcity,i (t) =
7 More

oi,r ∈Oi

|Oi |

(4.6)

components can be easily integrated. Examples are quality, relevance to the system, and
diversity of resources.

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

95

The relative scarcity of a resource can be determined by Equation (4.7), where ri,r
are the requests for resource r by user i, and oi,r are the offers of resource r by user i.
ηr =

∑i ri,r
∑i oi,r

(4.7)

An s-shaped scarcity function is used to normalize the scarcity value to the interval
[0, 1]. Equation (4.8) defines this function, where κ T is a threshold value that can
be set to specify the amount from which a resource is considered is scarce. This
approach allows for modeling flexibility by allowing to set different thresholds for
different types of resources.

ψr =

⎧
⎪
⎪
⎪
⎨0
1

2
⎪
⎪
⎪
⎩1

for ηr < 1

· sin( κπT · ηr − 2κπT (κ T + 2)) + 12

for 1 ≤ ηr ≤ κ T

(4.8)

for ηr > κ T

Trust-based Allocation Matching
Given the resource requests from other users, the selection process of a user has
to be modeled. In this case, trust-based resource allocation is used. Each user has
a certain (resource-specific) trust value with every other user, which depends on
previous experiences between the users (granted and fulfilled resource requests
leading to higher trust).
Initially, in a real-world Social Cloud users can categorize their friends with respect
to trust values. In the simulation, the initial assignment of trust values is based on
the network structure. Directly connected users are randomly assigned to trust
scores according to ϑ ∈ [0.5, 1.0]. For not directly connected users the initial assignment is based on trust transitivity. As argued by Golbeck (2005), trust is to some
degree transitive, meaning that the direct trust scores on the path of an indirect relationship can be used as a proxy for indirect trust. However, the exact degree of
transitivity is not clear, and the survey results in Section 3.3 indicate that trust to a
friend-of-friend is considerably lower than trust for users with a direct social connection. For this reason, and as repeated multiplication for indirect relationships
of a higher degree quickly decreases the corresponding trust value, the case study
assumes that trust is transitive for friend-of-friend relationships, and that the initial
trust value for connections involving more than one friend in between the users is
set to zero.

96

Designing Incentive Schemes and Co-operative Infrastructures

The trust score dynamically changes based on the rating that each user gets from
other users for the provisioning of resources or services. Similar to the trust-based
recommendation system described in (Walter et al., 2008), users are able to provide feedback about the quality of resource provision, where χr,j,i (t) denotes the
feedback of user j about user i for resource type r.
The update of trust values is done through following formula:
⎧
⎨ w · ϑ ( t ) + (1 − w ) · χ ( t )
t
t
i,j,r
r,j,i
ϑi,j,r (t + 1) =
⎩ (1 − w t ) · ϑ ( t ) + w t · χ ( t )
i,j,r

r,j,i

for χr,j,i (t) ≥ 0.5
for χr,j,i (t) < 0.5

(4.9)

In this formula, wt is the weighting factor that determines the smoothness of
the moving average, i.e., how fast new information (positive or negative feedback) χr,j,i (t) ∈ [0, 1] is taken into account (where user j rates user i). For values
wt ∈ [0.5, 1], this update scheme propagates negative effects fast and positive effects slow, which is representative for the property of trust (Walter et al., 2008), and
is also aligned with trust aspects in Social Clouds (Caton et al., 2012). In the simulation, wt = 0.75 is used as baseline value for this reason. To study the sensitivity
of the system with respect to this update rule, an alternative trust update scenario
(ATU) with wt = 0.25 is studied as well, in which case positive feedback effects are
incorporated fast whereas negative effects slowly.
For every resource r, the potential matches are shown to user i in a list sorted according to the trust values. The probability of selecting one of the requests of the
list is simulated through the Click through rates of Google (Hearne, 2006), as a
proxy for real choice behavior. In other words, this probability function assigns
a higher probability for matches on top of the list, as these are chosen more likely
than lower-ranked matches. Additionally, users are modeled to have a certain minimum trust threshold for sharing, ϑmin , that indicates the minimum trust score that
has to exist between users. In the simulation, ϑmin = 0.3 is chosen to reflect that
even in case of transitive trust, users might not be willing to share with users they
do not know (e.g. friends of friends) or have low trust ratings.8

8 Due

to the other assumptions and the relatively low level, the simulation results do not seem to
be sensitive towards the chosen value.

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

97

4.2.3. Evaluation of Dynamic Effects
The evaluation of the incentive scheme consists of two parts. First, the effect of the
participation constraints on certain performance indicators of the Social Cloud is
studied. In addition, the dependency of underlying user type distributions on the
results is investigated. Hence, the evaluation aims at addressing following questions considering the usefulness of a trading constraint for the resource exchange
platform:
• Effects of trading constraint: What is the effect of introducing the trading constraint on the average users’ utility? How does the trading constraint
change the number of successfully shared resources?
• Consequence for different user types: To what degree are different user
types affected by the trading constraint?

Simulation Specifics
For the case study, a social network with small-world properties is used as network
topology (using a connection rewiring probability of 40%, see Watts and Strogatz
(1998)). To study the effect of network size, networks of 20, 50, 100, 200 and 500
users are compared. Depending on the size, each user has a direct connection to
roughly one quarter of the other users, which also reflects certain small-world network properties.9 As a baseline, κ1 = κ2 = 0.5 is used in the simulation. For the trust
calculation, the average feedback given by users is χr,j,i (t) = 0.7 with a standard
deviation of 0.3 to account for changing quality in the requests. The scenario considered 3 resource types, a maximum resource amount of 10, as well as a scarcity
value of κ T = 5. All users are initialized with the same contribution score, which
changes over time according to the individual contributions of the users.
In all scenarios, 10 repetitions of the setting were made to minimize the potential
effects of initialization and random number effects. Results were averaged over
the 10 repetitions. Each scenario was implemented as a discrete event simulation
with 1000 simulation periods. In each period, the users decided on the amount of
resources they offer (based on the utility function in Section 4.2.2), and the matching is performed. In order to avoid the start-up problem and potentially skewed
9 For

example, in a network with 500 users, each user is directly connected to 100 other users.

98

Designing Incentive Schemes and Co-operative Infrastructures

results due to initialization effects (Law, 2007, 508ff.), only the last 900 periods are
used for the evaluation.

Effects of the Trading Constraint on the System

AverageUtilityperUser

While the introduction of a trading constraint as shown in the Section 4.2.2 was
proven to increase participation and reduce free-riding in P2P systems (see Ranganathan et al. (2004)), its usefulness in social collaborative scenarios is not clear.
Therefore, the effects of the trading constraint on the system are studied first. This
is followed by a more detailed analysis of the trading constraint on the different
user types.
1.2
1.1
1
0.9
0.8
0.7
0.6
20

50

100
NetworkSize

200

500

NoTradingConstraint

TradingConstraint

NoTradingConstraintͲATU

TradingConstraintͲATU

Figure 4.1.: Average User Utility

Figure 4.1 shows the average utility per user for different network sizes for the
baseline trust update as well as the alternative scenario where negative feedback
propagates slowly (ATU). Two effects can be observed. First, the average utility
increases with the network size. This is intuitive, as a larger network (and more
connections to other users) provides each user with more possibilities to share resources. That is, the larger a network and the more connections a user has, the
higher the likelihood of an accepted request. Second, the average utility per user
is lower in the scenario with the active trading constraint. This could also be expected, as the trading constraint yields lower contribution scores for free-riders,
thus indirectly lowering their utility. As the trading constraint limits the number

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

99

of users each user can share resources with10 , less requests are accepted, which
lowers the utility of users. Additionally, both effects are not sensitive to the trust
update mechanism, i.e., whether negative effects are propagated slowly in the trust
score or not. This can be seen by the qualitatively similar average user utilities for
the two different trust update mechanisms in Figure 4.1. This is an indication that
the qualitative effects are related to other factors such as the user types and their
behavior, rather than the trust calculation.
Looking at additional simulation results also yields interesting observations. Figure 4.2a shows the effect of the trading constraint on the probability that a resource
requests is successfully granted (i.e., matched with a suitable offer). Clearly, due
to similar reasons as for the decrease in average utility, the percentage of granted
resource requests is decreased if the trading constraint is introduced. As users, on
average, have a smaller set of other users they can potentially share resources with,
the probability that a request is successfully matched is decreased, on average.

0.5

RatioMatchesperOffer

PercentageGrantedRequests

Considering the trust update, Figure 4.2a shows that especially smaller networks
are affected by the trust update mechanism. If negative feedback affects the trust
score to a high degree, it is likelier that users affected by this fall below the necessary trust threshold needed for a successful transaction. This leads to the lower
percentage of successful requests compared to the alternative scenario (ATU) with
slow negative propagation.

0.4
0.3
0.2
0.1
0
20

50

100
NetworkSize

200

500

0.6
0.5
0.4
0.3
0.2
0.1
0
20

50

100
NetworkSize

200

500

NoTradingConstraint

TradingConstraint

NoTradingConstraint

TradingConstraint

NoTradingConstraingͲATU

TradingConstraintͲATU

NoTradingConstraintͲATU

TradingConstraintͲATU

(a) Percentage of Granted Resource Requests

(b) Ratio Matches per Offer

Figure 4.2.: Average Offers and Matches per Offer

Furthermore, while the simulation results reveal that the average amount of offers
per user does not significantly change with the network size, the ratio of successful
10 I.e., the subset of users each user can share with is strictly included in the set of users in the setting

without the trading constraint.

100

Designing Incentive Schemes and Co-operative Infrastructures

matches per offer significantly increases with the network size (see Figure 4.2b).
This can be explained by the fact that the more connections each user has, the
higher the potential number of other users that could be interested in a specific
offer. Again, due to the contribution score which reduces the potential set of sharing users, the scenario with the active trading constraint yields a lower match per
offer ratio. Similarly, the scenario with fast negative trust propagation yields lower
ratios as well.

Effect of Trading Constraint on User Types
The previous results are expected, due to the design of the trading constraint. However, it is at first not clear how different user types are affected by the trading constraint. The contribution score aims to provide incentives to share resources, as
selfish users with a low contribution score will not be able to request resources
from users with a higher contribution score. To study the second question stated
at the beginning of this case study, the ratio of utility per user in the active trading
constraint scenario is compared with the baseline scenario without trading constraint. Figure 4.3 shows the baseline trust update scenario, which yields several
interesting results.11

RelativeUtilitycomparedtoNoTrading
Constraint

1.2
1
0.8

Type1:Selfish
Type2:Leontief

0.6

Type3:PerfectSubstitutes
Type4:WeakSelfish

0.4

Type5:WeakLeontief
Type6:WeakPerfectSubstitutes

0.2
0
20

50

100
200
NetworkSize

500

Figure 4.3.: Relative Change in User Utility through the Trading Constraint

11 The

alternative trust update scenario yields very similar results and is therefore omitted.

4.2. CASE STUDY: DESIGNING INCENTIVE SCHEMES

101

First, as observed earlier, the average utility of some user types increases with the
network size, due to the larger pool of potential resource sharing partners. Second, the trading constraint discriminates users of different utility types. Having a
closer look, it can be seen that user types 1 (selfish) and 4 (weak selfish) are particularly vulnerable to the trading constraint, and their utility even decreases with
the network size. Indeed, these two user types represent selfish users who are only
interested in their own resource consumption. Due to the contribution score, their
set of potential sharing partners is reduced, which leads to less granted requests
in the scenario with active trading constraint. Hence, the simulation confirms that
selfish users (free-riders) are punished by the incentive scheme.
In contrast to the selfish users, other utility types perform better, especially with
increasing network sizes. The utility of user types 3 (perfect substitutes), 5 (weak
Leontief) and 6 (weak perfect substitutes, see Table 4.2) is close to the utility in the
scenario without trading constraint, especially with increasing network sizes. This
confirms that non-selfish users have less negative effects due to the contribution
score. In particular, user type 5 performs as well as in the baseline scenario for
larger network sizes, which means they are not affected at all by the trading constraint. In addition, the utility of user type 2 (Leontief) is not affected by the trading
constraint.12

Effect of User Type Distribution
The previous results are based on the distribution of utility types as given in (Andreoni and Miller, 2002). As this is a rather specific assumption which potentially
influences the results considerably, this section studies two different types of utility distributions in order to compare the validity of the previous findings. The first
user type distribution considers a larger percentage of social user types, which is
modeled as 70% of type 6 and only 30% selfish users of type 1. The second distribution considers the opposite, 70% selfish users (type 1) and only 30% of more
social users (type 6), based on empirical findings that many online and P2P networks exhibit a large number of non-contributing users (e.g., Adar and Huberman
(2000)).
12 This

specific result, however, should not be overemphasized, as this is a very specific utility function and in this simple scenario, users only apply a simple optimization function which ultimately determines their utility (in fact, the minimum in the Leontief utility function is responsible for the unchanged utility).

102

Designing Incentive Schemes and Co-operative Infrastructures

Table 4.3.: Incentive scheme with different user type distributions. Numbers in brackets are
average utilities without and with active trading constraint.
Users

Social Scenario
Average Relative Util- Relative UtilOffers ity Type 1
ity Type 6

Freerider Scenario
Average Relative Util- Relative UtilOffers ity Type 1
ity Type 6

20
50
100
200
500

1.38
1.39
1.42
1.40
1.46

0.65
0.73
0.66
0.59
0.65

0.87 (1.15,1.00)
0.80 (1.25,1.00)
0.77 (1.29,1.00)
0.76 (1.31,1.00)
0.75 (1.33,1.00)

0.87 (1.36,1.18)
0.94 (1.46,1.37)
0.96 (1.51,1.44)
0.97 (1.53,1.48)
0.98 (1.55,1.52)

0.92 (1.09,1.00)
0.86 (1.16,1.00)
0.84 (1.19,1.00)
0.82 (1.21,1.00)
0.82 (1.22,1.00)

0.92 (1.28,1.18)
0.93 (1.36,1.26)
1.01 (1.40,1.42)
1.04 (1.42,1.47)
1.06 (1.43,1.52)

Table 4.3 shows the relative utilities of the user types as well as the average number of offers per user for the two scenarios. The more social scenario with the
larger number of sharing users yields almost twice the average number of offered
resources per user, and also a higher average utility than the baseline scenario. In
addition, the relative decrease in utility for selfish users is higher in the social scenario and lower in the free-rider scenario. This is due to the higher utility for selfish
users in the social scenario without trading constraint, as they are able to benefit
from the higher number of offered resources by altruistic users and thus have a
higher utility in that case. Once the trading constraint is active, though, the average utilities of both types are same for both scenarios (based on the fact that in
neither scenario their resource requests are fulfilled).
For the type 6 users, an additional interesting effect can be observed. In the freeriding scenario, the relative utility is sometimes even higher with the trading constraint than without, whereas for the social scenario the relative utility is smaller
with the trading constraint. This can be explained by the observation that in the
free-riding scenario without trading constraints, users of type 6 have to compete
with selfish users, hence potentially lowering their chance of getting a resource
request granted. With the trading constraint and with a higher number of users
in the system, they only have to compete with other users of the same type, thus
increasing their chances to receive resources.
Overall, the results for the different user type distributions show that, in general,
the effect of the trading constraint is qualitatively the same (punishing free-riders).
In certain scenarios it can even be beneficial for more altruistic users to introduce
the trading constraint.

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

103

4.2.4. Summary
Summarizing the findings, the incentive scheme with the trading constraint discriminates between different user types. While selfish types are punished by the
constraint and their utility is significantly lowered, non-selfish users often perform
almost as well as, and sometimes even better, as the baseline scenario. Hence, this
simple case study outlines the usefulness of using simulation as tool in the design
process for Social Clouds. It both confirms the expected behavior of the incentive
scheme, and at the same time discovers interesting relationships concerning different user type distributions.
The case study can be augmented by a more detailed analysis to determine the sensitivity of the results on the simulation assumptions. Further sensitivity analyses
include the effect of different network structures (random, small-world, etc.), additional parameters in the contribution score (e.g., quality of the provided resources),
and different values for a trust threshold, below which resources are not granted to
a requesting user.

4.3. Case Study: Co-operative Infrastructures
Computational social resource sharing platforms such as Social Clouds aim at facilitating the exchange of resources among users by providing means to register as a
user, advertise or request resources, user lookup queries, communication utilities,
allocation algorithms, and so on. All these tasks require infrastructure resources to
store and use data, and the platform owner/provider has to decide how to acquire
these infrastructure resources. Two of the most common means are to either invest
in dedicated computational infrastructure resources on which the platform runs,
or to use third-party Cloud resources to host the platform. Both approaches incur
costs for the infrastructure and its maintenance.
Yet, a currently underrepresented approach to provide infrastructure resources to
host a platform is to make use of the computational resources available to the platform users in a co-operative resource provisioning model. Especially when users
are willing to share and exchange computational resources between other users,
they might be willing to share a certain part of these resources with the platform
itself. Such platforms already exist, for example cloud storage solutions that use

104

Designing Incentive Schemes and Co-operative Infrastructures

storage space provided by users, which can increase their allowed space by contributing to the platform.13
Such an approach is perfectly suitable for Social Clouds. As the platform requires
coordination mechanisms to facilitate their basic functionality (user management,
resource allocation, etc.), an underlying computational infrastructure must be provided. Supporting a distributed computing platform requires an initial investment,
advertising, or the introduction of fees. In a social context, this might be undesirable or counterproductive. For example, Shampanier et al. (2007) showed that users
will not even pay negligible fees when free alternatives are available.
Following this reasoning, the second case study proposes a novel co-operative infrastructure on which socially oriented exchange platforms, such as a Social Cloud
and potentially even social network platforms, can be supported, without incurring
the overhead or costs of provisioning dedicated infrastructure resources. Instead of
using dedicated, out-sourced, or third-party infrastructure, the platform is hosted
upon the computational resources it manages.
Such a co-operative scenario is particularly fitting for social and volunteer computing, as users are providers as well as beneficiaries of the platform. In this setting,
resources are defined as the computational capabilities needed by the collaborative
platform to function, which in the simplest case includes: computational power
and storage. This can also be extended to services like (distributed) databases and
P2P overlay networks. Although aspects like reliability, redundancy, and technical
instrumentation are important, these are not focus of this section. Instead, this section focuses on the economic methods to address system availability, redundancy,
and scalability.
The section is an extended version of Haas et al. (2013) and is structured as follows. Section 4.3.1 defines the notion of co-operative infrastructures and gives an
overview on related work and similar concepts. Section 4.3.2 introduces and describes the economic model behind co-operative infrastructures, and Section 4.3.3
evaluates the economic model with respect to certain performance metrics. Section
4.3.4 concludes with a summary and outlook on model extensions.

13 See

http://www.symform.com – last accessed May 2014

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

105

4.3.1. Definition of Co-operative Infrastructures and Related Work
Before the case study can consider the effects of contribution schemes for cooperative infrastructures, a general definition of this concept is needed to provide a
basis for the subsequent model. The unifying characteristic of practically all types
of co-operative organizations, such as credit unions, grocery shops, or infrastructures, is the fact that it is owned by a group of people. Depending on the type of
co-operative, not only the ownership but also the management of the organization
is shared by its users. These characteristics are the basis of the following definition
of co-operatively provided infrastructures:
Definition 5 ( Co-operative Infrastructure ). A co-operative infrastructure is a scalable
computing platform where all (computational) resources constituting the platform’s infrastructure, as well as those made available over the platform, are owned and/or managed by
its users.
Co-operatives can provide specific advantages over other methods, especially over
third party solutions. For example, it is difficult to quantify the platform’s required
quality of service. Consequently, commercial entities may be incentivized to act opportunistically, i.e., maximize profit, which might not be in the users’ best interest
(Spear, 2000). Therefore, trust in the platform is critical, and inherent in the concept
of collaboration and co-ownership in a co-operative.

User View on Co-operative Infrastructures
From the viewpoint of a user within a Social Cloud which utilizes co-operative
infrastructures, the decision to share resources involves two levels as shown in
Figure 4.4. Given that the user has resources available for sharing or exchange,
they can either share them directly with other users, or contribute the resources
to the co-operative infrastructure. These decisions are naturally intertwined, as
making the resources available to either the infrastructure or to other users implies
that the resources are not available for other purposes anymore. In other words,
the contribution of a user to the sharing platform depends on the contribution to
the co-op infrastructure, and vice versa.
An important implication of this viewpoint is that models for sharing resources
within the resource exchange environment should incorporate the possibility to

106

Designing Incentive Schemes and Co-operative Infrastructures

Share resources
with other users

Contribute resources
to infrastructure

Exchange
Level

Infrastructure
Level

Figure 4.4.: Levels of Sharing for Platforms with Co-operative Infrastructures

provide resource to the infrastructure, thus blocking these resources from sharing
with other users. However, this is not an easy task for several reasons. On one
hand, the decision of users to either use their available resources for sharing with
other users or for the infrastructure might be highly dynamic, and might depend on
factors such as current requests of certain friends or the overall contribution to the
co-op platform. On the other hand, the individual motivations to contribute to the
infrastructure might be different compared to sharing the resources with friends
(e.g., reciprocity considerations), and depend on the respective incentive scheme.
Hence, an integrated model is likely to be too complex for the given case and the
two levels are considered independently. Yet, the independent models can include
parameters that approximate the contribution to the other level. For example, in the
co-op model presented in this chapter, the parameter σi represents the percentage
of resources blocked for other uses, such as own consumption or sharing with other
users.

Related Work
Co-operative (co-op) business models are prevalent in a range of domains including grocery stores, credit/banking unions, health care and even housing. The
core premise of a co-op model is distributing ownership, management and profitsharing across its members. Depending on the scenario, co-ops can be formed between different groups of individuals and for many different reasons. Consumers,
employees, producers, and residents are all motivated in different scenarios to form
a co-op business. From an economic point of view, Porter and Scully (1987) investigate the formation of co-ops by comparing their efficiency to other organization

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

107

forms. They compared different efficiency metrics (price, scale and managerial
efficiency) and found that non-cooperative firms tend to be more efficient than coops. These metrics are, however, not applicable in this scenario as they are pricebased, which implies that other metrics are needed in case of non-monetary Social
Clouds.
In computer science, a co-op metacomputer was first introduced by Cime and
Marzullo (1999). In this system, contribution is voluntary but intended to be mutually beneficial for users. More recently, in Grid and Cloud computing, federations have been created that span multiple institutions allowing resources to be
shared amongst members of a virtual organization (VO). In these domains, novel
co-op approaches to resource management and task allocation such as the Community Scheduler Framework (CSF) (Xiaohui et al., 2006) and the DRIVE metascheduler (Chard and Bubendorfer, 2008) have been proposed. Both CSF and DRIVE
utilize community contributed resources for core metascheduling tasks such as
resource allocation. Considering the feasibility of that approach, Xu and Yung
(2010) showed that secure and privacy preserving auction protocols can be used
in a co-operative metascheduling architecture to conduct trustworthy resource allocations on potentially untrusted resources. Unlike the Social Cloud model, these
metaschedulers are designed to support job submission on a large scale, comparably static, Grid environment where providers have huge resource pools and can
provide explicit availability guarantees.
The concept of a co-op is implicit in the definition of a Social Cloud, as members
benefit directly from resource sharing with one another. A Social Cloud differs from
many other resource sharing architectures as it leverages the relationships defined
in a social network. However, Social Clouds and therefore the co-op infrastructure
are not free from risk, as relationships can change over time. The implications for
security in dynamic Social Cloud environments have been investigated by Xu and
Yung (2010). Socially based systems are also becoming increasingly common in
both academic and commercial applications. For example, social networking principles are commonly employed for coordinating ad hoc research communities such
as the previously mentioned myExperiment.org (Roure et al., 2009), nanoHUB.org
(Klimeck et al., 2008), and GlobusOnline14 (Foster, 2011). Commercial applications
such as FriendStore15 (Tran et al., 2008) offer distributed file storage provided by a
14 https://www.globus.org/

– last accessed May 2014
– last accessed May 2014

15 http://friendstore.news.cs.nyu.edu/

108

Designing Incentive Schemes and Co-operative Infrastructures

user’s friends. Diaspora16 is a first step towards creating a Facebook-like social network platform relying on content resources (e.g., images) provided by its members
(rather than dedicated resources). The model proposed in this case study builds
upon this idea, by also capturing the computational resources needed to run the
system, providing a platform free of dedicated centralized servers.
The concept of co-operative infrastructures is conceptually similar to several other
sharing scenarios. For example, in public goods scenarios users contribute parts of
their resources to a public good which is accessible to other users (e.g., paying taxes
where the projects funded by tax money can be subsumed as public goods). Another example is the sharing of knowledge in related communities. In both cases,
the contribution of each user affects the other users as they are able to consume the
public good. As mentioned in Section 3.2, the “Tragedy of the Commons” problem can be observed in many related systems, which states that in public goods
scenarios a single user’s incentives to contribute might be inversely affected by
the contributions of other users. Hence, this issue also has to be addressed in a
co-operative setting. In the following case study, this is done through certain contribution schemes by which the users contribute to the infrastructure.
Whereas this case study focuses on the economic model of contribution schemes,
similar systems have been studied from a conceptual and technical perspective. Babaoglu et al. (2012) consider the creation of P2P Infrastructure Clouds in
which the (potentially unreliable) cloud resources are provided by the participants
of the cloud. Their prototype implementation focuses on aspects such as selforganization and robustness to failures. In another model, Khan et al. (2013) study
infrastructure clouds based on community networks. Their evaluation focuses on
technical aspects such as response time and its dependency on the heterogeneity
of provided resources. Both approaches, however, focus on technical issues and do
not consider resource contribution schemes.

4.3.2. Economic Model
Previous literature on co-operative models focuses either on the organizational
level (Porter and Scully, 1987; Spear, 2000), or the individual member level (Sexton,
1986). As no existing organizational data is available for this study, the approach
of Sexton (1986) is applied which looks at individuals incentives to participate in
16 https://joindiaspora.com/

– last accessed May 2014

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

109

the co-op.17 To do this, the following aspects need to be specified: 1) the users
and their characteristics (e.g., endowment, utility function, etc.) (Sexton, 1986); 2)
the resource requirements of the distributed system (Cime and Marzullo, 1999), in
this case the Social Cloud platform; and 3) the contribution schemes through which
users are encouraged to join the co-op and provide resources (Sexton, 1986). Each
of these aspects are elaborated on in the following subsections to create a co-op
model that is adequate for the considered scenario. As discussed, only computational infrastructure resources are considered in the co-op model, not software
requirements such as databases.

User Model
The basic user model was presented in Section 4.1. In particular, the six different
utility functions as given in Table 4.2 are used to model six different types of users.
For this case study, the availability of users is additionally considered. Users are
modeled as having a rate of availability αi , in which their resources are available
to the Social Cloud. This aspect constitutes the notion of quality of (a contributed)
service (QoS), as availability is a standard QoS measure.
As Social Cloud implementations are still in their prototype phase, no traces of user
behavior within the context of a Social Cloud exist. Consequently, the performance
of a Social Cloud cannot be analyzed based upon production user data. However,
to develop a co-op model it is assumed that user behavior within a Social Cloud
is similar to behavior patterns in Volunteer Computing, as both rely on ad-hoc
user communities donating spare computing capacity. Therefore, to model users
the SETI@home (Anderson, 2004) host availability traces for the years 2007-09 is
selected, which contain information for over 220,000 hosts.18 These traces have
been the subject of many analytical studies (e.g., Anderson and Fedak (2006); Javadi
et al. (2009, 2011)), motivating their selection, as they provide a solid foundation to
base a Social Cloud context upon. Statistical analysis of the data in (Javadi et al.,
2011) reveals that 21% of the monitored hosts exhibit statistical independence in
17 It

should be noted that this is only one possible approach. The fact that economic actions are influenced by the social relations of a user has long been recognized (see e.g. Granovetter (1985)).
Yet, it can be argued that these approaches (taking social embeddedness into account) are especially helpful when applied to (economic) network analysis based on observed data. However,
as the actions of individual users are simulated without existing data, the model relies on specific
behaviors instrumented through theoretically grounded and context fitting utility functions.
18 Available: http://fta.inria.fr – last accessed May 2014

110

Designing Incentive Schemes and Co-operative Infrastructures

their availability. The authors also provide probability distributions for these hosts
as well as for several clusters that can be distinguished within these 21% of hosts.
These clusters are used to define different user types in the co-op model. Based on
these results, Javadi et al. (2011) deduced a distribution of the length of availability
intervals, which is used to simulate the availability of users. Table 4.4 summarizes
the statistics of the six clusters, providing their relative size, average availability
and the distribution functions of the interval lengths.19
Table 4.4.: List of User Cluster based on Javadi et al. (2011)
Cluster

Relative Size

Avg. Availability

Avail. Interval Dist.

Unavail. Interval Dist.

1
2
3
4
5
6

0.042
0.108
0.658
0.004
0.094
0.094

83.9%
82.2%
26.0%
91.7%
72.3%
56.7%

Gamma(0.289,311.711)
Gamma(0.340,152.216)
Weibull(0.431,1.682)
Gamma(0.347,371.622)
Gamma(0.342,89.223)
Gamma(0.357,43.652)

Hyper-Exponential
Hyper-Exponential
Hyper-Exponential
Hyper-Exponential
Hyper-Exponential
Hyper-Exponential

While the user types in Table 4.2 are found in experimental games, one might ask
if this is a realistic setting for the case of co-operative infrastructures. Hence, this
setting is also compared with another user type distribution which considers sharing in P2P networks. The basis for this distribution is obtained from Adar and
Huberman (2000), who present a study on free-riding in Gnutella, a then-popular
P2P platform, and shows that in this platform approximately 70% of the users do
not share at all. Therefore, another scenario is introduced which captures this user
type distribution in which 70% of users have a selfish utility function of type 1, and
the remaining 30% share to some degree (according to utility function type 6). The
second scenario is used to study the effects of different user type contributions on
the applicability of the considered contribution schemes.

Modeling System Requirements
Determining the feasibility of users providing the computational infrastructure of
a platform requires knowledge of the platform’s resource requirements (i.e., the resources required to guarantee the functionality of the platform). In general, system
requirements R (n) can be modeled as R (n) = l1 f (n) + l2 where l1 f (n) specifies
the increase in resources based on the number of users (n) and l2 is the minimum
19 The

specifics of the unavailability interval distributions can be found in Javadi et al. (2011).

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

111

amount of infrastructure resources needed for the platform. This is a reasonable
base model as the computational requirements for necessary platform tasks generally increases with the number of users (and transactions), while some base functionality must also be provided at all times (e.g., database services).
A realistic model of required resources is essential for the co-op model and the
evaluation results. Hence, the system requirements function is modeled based on
certain assumptions as well as empirical observations from the Social Cloud prototype presented and evaluated in Chard et al. (2010). They studied several aspects
of the deployed prototype, such as resource (CPU, memory) usage, on which the
requirements function is based. In abstract terms, the system requirements are
calculated for one time period. In this time period, each of the n users participates in y transactions which are typical for the platform. For example, this could
be the computation of an auction or preference-based allocation for a particular
resource. The reverse second-price-sealed-bid auction as implemented in Chard
et al. (2010) is considered as a proxy for the transaction. This mechanism is known
to be resource-demanding and has high resource requirements, thus representing
the necessary computational power needed to host a Social Cloud and ensure its
functionality.
Every transaction requires a certain amount of resources for execution. From Chard
et al. (2010) it can be empirically derived that an average
computer (represented as
√
3

2

the average resource amount, ω̄) requires f (n) = ω̄ γn resources for one transaction.20 As it is unrealistic that all transactions in one time period have to run simultaneously, the parameter δ denotes the percentage of transactions that require
resources at the same time. If ωmin denotes the minimum number of resources that
have to be provided regardless of supporting basic functionality, then the system
requirements function (SR1 ) can be expressed as:

√
3
SR1 : R (n) = δ ∗ y ∗ n ∗ ω̄

n2
+ ωmin
γ

(4.10)

Equation (4.10) assumes that all users participate in every transaction. If for example, only half of the users, on average, participate in transactions then the system

exact function was found using curve fitting in Matlab, which yields γ = 933 and a goodness
of fit of 0.97. Note that Chard et al. (2010) provides values for up to 50 users, but the findings
strongly suggest that the general trend of the curve can be extrapolated for more users.

20 The

112

Designing Incentive Schemes and Co-operative Infrastructures

requirements function (SR2 ) will be:

3

SR2 : R (n) = δ ∗ y ∗ n ∗ ω̄

(n/2)2
+ ωmin
γ

(4.11)

Contribution Schemes
Considering the social context of the Social Cloud paradigm, it can be argued that
the goal of such a system should be to provide sufficient resources to run the platform, but at the same time not burden users with too much effort or high costs.
If ri is the resource contribution of user i to the co-op infrastructure, the following
optimization problem has to be solved from a system-wide perspective:
maxρi

∑Ui (ρi , σi )
i

n

s.t.

∑ ri ≥ R ( n )

i =1

r i = ρ i ∗ ωi

∀i ∈ 1, ..., n

σi + pρi = 1

∀i ∈ 1, ..., n

σi , ρi ≥ 0,

∀i ∈ 1, ..., n

(4.12)

In order to acquire the required infrastructure resources by the co-op’s members,
the following contribution schemes are considered:
• Enforced fixed contribution (EFC): Users are required to provide a certain
quantity of resources to the infrastructure as a membership requirement.
• Voluntary fixed contribution (VFC): Users choose to contribute a predefined
percentage of their own resources.
• Voluntary variable contribution (VVC): Users may freely choose what percentage of their endowment they are willing to contribute, i.e., no minimum
is prescribed.
Enforced Fixed Contribution For fixed contribution schemes, two alternatives are
possible: either the contribution can be fixed independent of the user’s resource
endowment, or the contribution can be proportional to the user’s endowment. The
first alternative is used by some popular file sharing systems, where a minimum

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

113

amount of resource sharing is required. However, such a scheme might exclude
users from the system which do not have the required resources available, or would
force some users to allocate most of their resources for the infrastructure. Hence,
the latter case is used to calculate the fixed percentages, because it is more considerate of users with low resource endowments.
In the proportional contribution scenario, let ρi∗ denote the percentage of user i’s
resource endowment that needs to be allocated to the co-op infrastructure. The
minimum necessary percentage per user, taking their average availabilities into
account, can then be calculated as:
ρi∗ = ρ∗min ∗ (1 − σi )
R (n)
ρ∗min =
∑i (1 − σi ) ∗ ωi ∗ αi

(4.13)
(4.14)

The calculation of the expected contributed resources depends on the assumptions
about the users, specifically σi and ωi . Without information about user types an
assumption has to be made, otherwise the expected or observed user type distributions can be used to estimate σi and ωi . Clearly, if the estimate of user types does
not reflect the true distribution, it will affect the overall amount of resources that
are contributed, and thus impact the availability of the system.
This calculation represents a static view of the system, as only average availabilities are taken into account. Dynamic effects such as the distribution of availability
and unavailability intervals might affect the overall performance of the system,
however, especially if these intervals are correlated between users (e.g., availability
patterns during the week compared to during the weekend). Hence, the feasibility of such a contribution scheme has to be studied via simulations to capture the
dynamic properties.
Voluntary Fixed Contribution In this scenario, the percentage ρ is prescribed by
the system according to some criteria, and users voluntarily contribute if ρ is
smaller than their individual optimum ρ∗ , otherwise they choose to not contribute
resources. This differentiates this scheme from EFC, as in the latter case a certain
contribution is mandatory.
Hence, the contribution of user i can be described as follows:

114

Designing Incentive Schemes and Co-operative Infrastructures

ρi =

⎧
⎨ρ

if ρ∗ ≥ ρ

⎩0

else

(4.15)

Although this scheme does not enforce contribution of users, it also significantly
depends on the choice of ρ. If it is set too low, more people will contribute, yet the
overall amount of contributed resources might not be high enough. On the other
hand, setting ρ too high potentially leads to less contributing users.
Voluntary Variable Contribution In this contribution scheme, users choose their
level of contribution based on their individual preferences for resource usage, considering for example altruistic motivation. This scheme addresses the key motivation of a Social Cloud, that is, users voluntarily choose to provide resources to the
platform.
Using the utility function described earlier (Eq. 4.2) and the resource constraint (Eq.
4.1), the optimization problem can be solved for user i. With the basic assumption
that Ui ≥ 0, the following result can be obtained:
∂Ui (σi , ρi )
=0
∂σi


1
1
β −1
−1
βΠi,s − βλp− β (ωi − Πi,s ) β−1 = 0
(Ui ) β
β


β −1
βΠi,s − βλp− β (ωi − Πi,s ) β−1 = 0

(4.16)
(4.17)
(4.18)
λ

Πi,s =
p

−β
1− β

−1
1− β
−1

+ λ 1− β

ωi

(4.19)

Hence, the relative contribution to the infrastructure depends on several factors: 1)
the importance of donating resources to the infrastructure (λ, the level of altruism),
2) the convexity of the utility function ( β), and 3) the relative price p for giving
resources to the infrastructure. In particular, the more expensive it is to allocate
resources to the infrastructure, compared to consuming them (i.e., p > 1), the lower
the users’ contribution to the infrastructure.

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

115

4.3.3. Evaluation of Contribution Schemes
Having developed a comprehensive economic co-op model, the focus of this evaluation is to investigate the following aspects of a co-op infrastructure for a Social
Cloud: 1) the average utility of users between the contribution schemes, and 2) the
effects of specific parameters, especially the relative price p to provide resources,
in the proposed contribution schemes. Additionally, two QoS factors of the platform are studied: system availability (as the percentage of time periods in which
the resource requirements are met) and the ratio of provided to required resources
(as a measure of redundancy). To study the effects of the number of users on the
results and the scalability of each approach, Social Clouds of different sizes (10
to 400 users) are considered. Note that these numbers are much smaller than the
hundreds of millions of users that are registered on Facebook, however individual
Social Clouds represent individual social network islands, and not the entire social
graph.

Simulation Specifics
The general simulation framework is implemented as follows. In total 100 simulation runs per scenario are performed and the results averaged, in order to account
for possible simulation bias with respect to the initialization of users. At the beginning of a simulation run, users are created according to the clusters in Table 4.4.
Users might have different resource endowments available, hence their respective
resource endowments are drawn uniformly from the interval [5, 15], and the average resource endowment is given by ω̄ = 10. As a baseline, the minimum amount
of resources required to host the Social Cloud is assumed to be half of an average
user’s resource endowment (i.e., 50% computation power of an average computer).
For the enforced fixed contribution scenario, ρi∗ is calculated according to the equations in Section 4.3.2, and σi is calculated based on the type of user i.
All scenarios include 10,000 simulated time periods. In the first period, it is determined whether users are initially available (or not) based on the average availability of the cluster, and the length of the first (un)availability interval is drawn according to the respective distribution. In every following time period, the amount
of resources provided for the infrastructure is calculated, and if an (un)availability
interval ends, the availability of a user is switched and the next interval is determined. Considering the startup problem (Law, 2007), the simulation requires ap-

116

Designing Incentive Schemes and Co-operative Infrastructures

proximately 50-500 time periods to reach a steady state, so only the results of the
time periods 500-10,000 are considered for the evaluation.
Two different system resource requirements functions, SR, are considered for each
contribution scheme in order to study the robustness of the findings. First, the
worst case requirements SR1 where R (n) is set according to Equation (4.10) with
δ = 1, y = 1, ω̄ = 10 and ωmin = 5. In this case, every user participates in every
transaction and all transactions have to be computed simultaneously, which reflects a peak scenario with very high system load. The values for ω̄ and ωmin are
determined because users’ endowments are drawn from [5, 15]. Second, the second
system requirements function SR2 represents an off-peak scenario with lower requirements, and is determined by Equation (4.11). For SR2 , only half of the users
participate in each transaction, and δ is set to 0.5. This is still a high average load,
yet acknowledges that the co-op infrastructure might be set up to capture average
requirements instead of peak loads.

Enforced Fixed Contribution
In this contribution scheme, each user has to provide a certain percentage of their
endowment to the system. For the evaluation, the actual resource contributions are
simulated when users provide exactly the minimum percentage (as calculated in
Equation 4.14), and also when they must provide 10%, 20% and 50% more than the
minimum percentage.
Figure 4.5 shows that the results of this contribution scheme depend on several factors: first, the minimum contribution percentage initially decreases with the number of users, but if the system size surpasses 50 users ρ∗min begins to increase again
(Figures 4.5c and 4.5d). Second, the number of provided resources is higher than
the amount of required resources (i.e., there is some level of redundancy, see Figures 4.5e and 4.5f). Third, the average availability of the system ranges from 80%
to 100%, for the optimal contribution ρ∗min (Figures 4.5a and 4.5b). This is mainly
due to the fact that the calculation of ρ∗min only considers average values of user
availability, which does not account for the dynamics of the system (such as multiple users being unavailable simultaneously). Furthermore, for worst case resource
requirements it begins to slightly decrease for larger systems. For average resource
requirements, the availability is slightly better for larger systems and remains constantly at 100%.

1.05
1.00
0.95
0.90
0.85
0.80
0.75
0.70

AvailabilityPercentage

AvailabilityPercentage

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

10

20

ʌ*

50
100
NumberofUsers

1.1ʌ*

1.2ʌ*

200

10
ʌ*

1.5ʌ*

0.3
0.2
0.1
0.0
20

ʌ*

50
100
NumberofUsers

1.1ʌ*

1.2ʌ*

200

400

RatioContributed/Required
Resources

4.0
3.0
2.0
1.0
0.0

ʌ*

20

1.1ʌ*

50
100
NumberofUsers
1.2ʌ*

200

1.1ʌ*

400

1.5ʌ*

(e) Provided vs. Required Resources, SR1

1.2ʌ*

200

400

1.5ʌ*

0.4
0.3
0.2
0.1
0.0
10

20

ʌ*

5.0

50
100
NumberofUsers

0.5

1.5ʌ*

(c) Contribution Percentages, SR1

10

ContributionPercentage

0.4

20

(b) Avg. System Availability, SR2

50
100
NumberofUsers

1.1ʌ*

1.2ʌ*

200

400

1.5ʌ*

(d) Contribution Percentages, SR2
RatioContributed/Required
Resources

ContributionPercentage

0.5

10

1.05
1.00
0.95
0.90
0.85
0.80
0.75
0.70

400

(a) Avg. System Availability, SR1

117

8.0
6.0
4.0
2.0
0.0
10

ʌ*

20

1.1ʌ*

50
100
NumberofUsers
1.2ʌ*

200

400

1.5ʌ*

(f) Provided vs. Required Resources, SR2

Figure 4.5.: Simulation Results for Enforced Fixed Contribution

Two issues can be observed with this contribution scheme. On one hand, if the minimum contribution percentage is increased to account for dynamic (un)availability,
the average availability of the system increases, reaching values very close or equal
to 1 (Figures 4.5a and 4.5b). On the other hand, the system is also scalable as the
average system availability increases with the number of users. This can be explained by the fact that the more users contribute, the larger the pooling effects
of (un)available users, which leads to a “smoothing” of provided resources and
a higher probability that the system resource requirements are met at any point
in time. Furthermore, the percentage of time periods where the system resource
requirements are actually met does not seem to depend on the specific resource

118

Designing Incentive Schemes and Co-operative Infrastructures

function as a comparison of Figures 4.5a and 4.5b shows. This is mainly due to the
fact that the (un)availability interval distributions determine the number of contributing users, and the requirements function mainly affects ρ∗min .
Overall, the results show that if an enforced fixed contribution scheme is used, it
is better to set the required ρ higher than the minimum ρ∗min . This leads to higher
system availability even for small increases of ρ because the larger individual user
contributions alleviate the unavailability of other users.

Voluntary Fixed Contribution
In this scenario certain users contribute a fixed percentage of resources if this percentage is lower or equal to their optimal percentage. The decision criterion can be
found in Equation 4.15 in Section 4.3.2. This scheme does not force users to contribute resources, as EFC does, yet the usability depends on the choice of the fixed
contribution percentage ρ as well as the user type distribution.
Figure 4.6 displays the simulation results where not all users have to contribute
to the co-op infrastructure. Several interesting findings can be observed. First, as
expected the percentage of users who contribute to the infrastructure decreases as
the required contribution amount increases (Figures 4.6c and 4.6d). Second, especially for high resource requirements the provided resources are lower than the required amount, therefore leading to a lower system availability as in EFC (Figures
4.6a, 4.6e as well as 4.6b and 4.6d). This indicates that the worst-case system requirements (SR1 ) increase disproportionally to the contributed resources for larger
system sizes (in this case, more then 50 users). For average resource requirements
(SR2 ), system availability is better and even can reach 100%, and the scalability
behavior improves as well.
In particular, the results also show that the system performance crucially depends
on the required contribution that is selected. Contribution levels between ρ = 0.2
and ρ = 0.3 seem to be most promising for the given scenario.

Voluntary Variable Contribution
The variable voluntary scheme in Section 4.3.2 allows each user to calculate their
optimal individual contribution for infrastructure resources, based on individual

1.0

AvailabilityPercentage

AvailabilityPercentage

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

0.8
0.6
0.4
0.2
0.0
10

20

ʌ=0.1

ʌ=0.2

50
100
NumberofUsers
ʌ=0.3

200

ʌ=0.4

1.2
1.0
0.8
0.6
0.4
0.2
0.0
10

400

0.10
0.05
0.00
10

20

ʌ=0.1

ʌ=0.2

50
100
NumberofUsers
ʌ=0.3

200

ʌ=0.4

400

RatioContributed/Required
Resources

1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0

ʌ=0.1

20

ʌ=0.2

50
100
NumberofUsers
ʌ=0.3

200

ʌ=0.4

400

ʌ=0.5

(e) Provided vs. Required Resources, SR1

ʌ=0.3

200

ʌ=0.4

400

ʌ=0.5

0.25
0.20
0.15
0.10
0.05
0.00
10

ʌ=0.5

(c) Contribution Percentages, SR1

10

PercentageContributingUsers

0.15

ʌ=0.2

50
100
NumberofUsers

(b) Avg. System Availability, SR2

ʌ=0.1

20

ʌ=0.2

50
100
NumberofUsers
ʌ=0.3

200

ʌ=0.4

400

ʌ=0.5

(d) Contribution Percentages, SR2
RatioContributed/Required
Resources

PercentageContributingUsers

(a) Avg. System Availability, SR1
0.20

20

ʌ=0.1

ʌ=0.5

0.25

119

3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
10

ʌ=0.1

20

ʌ=0.2

50
100
NumberofUsers
ʌ=0.3

200

ʌ=0.4

400

ʌ=0.5

(f) Provided vs. Required Resources, SR2

Figure 4.6.: Simulation Results for Voluntary Fixed Contribution

preferences, according to Equation (4.19). In the following simulations the relative
price of resource contribution p was also varied between 1.0 and 3.0 to capture
that providing resources might be comparably more expensive than using them or
leaving them idle.21 In order to study system behavior in a more realistic setting,
each user is assigned to a certain availability cluster and is given a particular type of
utility function, according to Table 4.2. The type of utility function also determines
p < 1 would mean resource provisioning is less expensive then leaving them idle, which
can be, e.g., achieved through subsidization. Although this might be an option to encourage
contribution to the co-op, it has to be determined where such subsidies would come from. This
is not the focus of this case study, hence the case p < 1 will not be considered in this evaluation.

21 Prices

Designing Incentive Schemes and Co-operative Infrastructures

1.2

AvailabilityPercentage

AvailabilityPercentage

120

1.0
0.8
0.6
0.4
0.2
0.0
10

20

p=1.0

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

0.00

p=1.25

p=1.5

p=2

RatioContributed/Required
Resources

p=1.0

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

400

p=3

(e) Provided vs. Required Resources, SR1

20

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

400

p=3

(b) Avg. System Availability, SR2
0.20
0.15
0.10
0.05
0.00
10

p=3

3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
20

0.0

400

(c) Contribution Percentages, SR1

10

0.2

p=1.0

ContributionPercentage

0.05

p=1.0

0.4

p=1.0

20

50
100
NumberofUsers

p=1.25

p=1.5

200

p=2

400

p=3

(d) Contribution Percentages, SR2
RatioContributed/Required
Resources

ContributionPercentage

0.10

200

0.6

p=3

0.15

50
100
NumberofUsers

0.8

10

(a) Avg. System Availability, SR1

20

1.0

400

0.20

10

1.2

7.0
6.0
5.0
4.0
3.0
2.0
1.0
0.0
10

p=1.0

20

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

400

p=3

(f) Provided vs. Required Resources, SR2

Figure 4.7.: Simulation Results for Voluntary Variable Contribution

the optimal contribution level, given the optimal condition based on the resource
endowment constraint.
Using the user type distribution from Andreoni and Miller (2002), the results
shown in Figure 4.7 are obtained. The simulation shows several interesting results.
On one hand, the intuition that higher relative contribution prices leads to fewer
contributions is confirmed in Figures 4.7c and 4.7d. The average contribution decreases with increasing relative prices for provisioning, a result which is independent of the system resource requirements. On the other hand, system availability

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

121

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0

AvailabilityPercentage

AvailabilityPercentage

as well as the average amount of contributed resources strongly depend on the relative prices that users bear for providing resources. If relative prices are too high
users choose to provide only a small number of resources, which leads to the decrease in system availability as shown in Figures 4.7a and 4.7b. This also affects the
scalability of the system; if prices for contribution are roughly similar to prices for
self-consumption, the contribution scheme is scalable with the number of users.

10

20

p=1.0

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

p=1.5

p=2

RatioContributed/Required
Resources

p=1.0

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

0.0

400

p=3

(e) Provided vs. Required Resources, SR1

20

p=1.0

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

400

p=3

(b) Avg. System Availability, SR2
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
10

p=3

1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
20

0.2

400

(c) Contribution Percentages, SR1

10

0.4

10

ContributionPercentage
p=1.25

200

0.6

p=1.0

20

50
100
NumberofUsers

p=1.25

p=1.5

200

p=2

400

p=3

(d) Contribution Percentages, SR2
RatioContributed/Required
Resources

ContributionPercentage

p=1.0

50
100
NumberofUsers

0.8

p=3

0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
20

1.0

400

(a) Avg. System Availability, SR1

10

1.2

3.0
2.5
2.0
1.5
1.0
0.5
0.0
10

p=1.0

20

p=1.25

50
100
NumberofUsers
p=1.5

200

p=2

400

p=3

(f) Provided vs. Required Resources, SR2

Figure 4.8.: Simulation Results for Voluntary Variable Contribution, mostly Selfish User
Types

122

Designing Incentive Schemes and Co-operative Infrastructures

Furthermore, on average, significantly more resources are provided than required,
leading to higher redundancy levels (Figures 4.7e and 4.7f). Along with the previous finding this indicates that the incentive-based scenario would be especially
vulnerable when only a few users contribute to the system. Although individually they might contribute a significant amount of resources, the system is more
dependent on these few contributors. In case of their unavailability the resource
requirements might not be met.
To study the dependency of the simulation results on the underlying distributions
of user types (and hence, utility functions), the previous results are now compared
with a more conservative distribution, as shown in Figure 4.8. As stated in Section
4.3.2, this distribution is obtained from observed P2P systems, where 70% of users
are purely selfish free-riders, and only 30% are willing to give resources to the
infrastructure (types 1 and 6 in Table 4.2, respectively).
In this scenario, based on the high number of free-riders, the general level of system
availability is much worse than in the previous case (Figures 4.8a and 4.8b). This is
especially evident at peak times where resource requirements are high, as the high
number of selfish users leads to a significant decrease in resource contribution. In
the case of moderate resource requirements, the system performs well when contribution is not expensive ( p ≈ 1) and for larger numbers of users. Moreover, both
the proportion of provided to required resources (Figures 4.8e and 4.8f), and the
average contribution per user (Figures 4.8c and 4.8d), are lower compared to the
results in Figure 4.7, which explains the reduced system availability. It is noteworthy that despite the conservative distribution of user types, the system performs
surprisingly well for moderate resource requirements and lower relative prices.

4.3.4. Discussion
Figure 4.9 shows a comparison between the average utility per user in the enforced
fixed, voluntary fixed and voluntary variable schemes and for both resource functions. As predicted, the average utility is higher when users are allowed to choose
their individual optimal level of contribution (rather than being forced into potentially suboptimal levels of contribution, as in the case of EFC and VFC).
Based on the previous evaluation results, there are some further observations
through which the feasibility and scalability of a co-op infrastructure can be evaluated. First, if a fixed contribution is enforced by the system, the (un)availability

4.3. CASE STUDY: CO-OPERATIVE INFRASTRUCTURES

123

3.7

AverageUserUtility

3.6
3.5
3.4
3.3
3.2
3.1
3
10

EFCSR_1

20

EFCSR_2

50
100
NumberofUsers
VFCSR_1

VFCSR_2

200

400

VVCSR_1

VVCSR_2

Figure 4.9.: Utility per User for Different Scenarios

characteristics of the users have to be taken into account when the necessary contribution percentage is calculated. In this case there is an inherent trade-off between
increasing the individual contribution ρi∗ and making the entire system perform
better or more reliably. That is, the QoS factor availability is directly dependent on
the chosen level of parameter ρi∗ . For example, an increase in ρi∗ significantly increases the system availability, but in this case individual contributions can be very
high for large system sizes (up to 47% of the endowment). Second, the results from
the voluntary variable contribution scheme indicate that even under worst case requirements, allowing users to select their contribution level the system resource
requirements are met in many time periods when relative contribution prices are
low (corresponding to high availability, as well as a rather high redundancy factor
as shown in Figures 4.7e and 4.7f). Furthermore, the average individual utility is
higher than in the fixed scheme. However, based on the underlying distributions
of user types, the system performance is much more dependent on the voluntary
contribution of single users.
Hence, the system designer has to address the following trade-off: either force
users to give a certain, fixed percentage of their endowment to the co-op infrastructure, e.g. as a condition for participation, or let users freely choose their contribution. In the latter case, the average utility of users as well as system availability/redundancy tends to be higher, but one must carefully study the user type
distribution to ensure that there are enough users willing to contribute. For practical settings, it might be beneficial to monitor the resource contributions and/or

124

Designing Incentive Schemes and Co-operative Infrastructures

determine the distribution of user types, and adjust the used contribution scheme
if the system availability level falls below a certain threshold.
There are several potential extensions of the co-operative infrastructure model that
can be pursued. First, the model can be augmented to account for additional resource types. Instead of only considering one resource type, real co-operative systems require several different types of resources to run, e.g., computational power
to calculate transactions, bandwidth to accommodate large file transfers, and storage to save databases and other information. In this case, the co-op infrastructure requires a certain amount of each resource type to be successfully deployed,
and over-provisioning of one resource cannot compensate underprovisioning of
another resource, leading to a combinatorial problem.
Another extension of the presented model is to analyze how specific incentives
(e.g., reputation mechanisms, compensation for resource provisioning, etc.) alter
the willingness of users to contribute. This is particularly interesting when combined with an integrated model that considers both sharing resources with other
users and contributing to the co-op infrastructure. Such a unified model would allow the design of tailored incentive schemes that take both aspects of contribution
into account.
Furthermore, as a co-op infrastructure induces both positive and negative external/network effects with a growing number of (contributing) users, another question is if there is an optimal size of a co-op infrastructure depending on the interplay of negative and positive effects. While this has been studied in context of P2P
music sharing networks (Asvanund et al., 2004), the adaption to co-op infrastructures remains an open issue. For example, the previous results show that for EFC,
the system availability level increases with a higher number of users which indicates that the co-op is more robust for larger number of users, whereas for VVC the
results depend on the relative price of resource provisioning. Finally, the ultimate
goal of a co-op model is to provide a sustainable infrastructure for its users. To
guarantee sustainability, concepts such as Key Performance Indicators (KPIs) can
be investigated and applied to the co-op model. Through the management of these
KPIs, the administrators of the co-op infrastructure could monitor, analyze, and
steer the system to ensure that the minimum service quality levels are fulfilled.

4.4. SUMMARY

125

4.4. Summary
This chapter presented two case studies how a simulation-based approach can be
leveraged in the design of incentive schemes for exchange and sharing platforms
such as Social Clouds. The focus of the first case study (Section 4.2) was the introduction of an incentive scheme in resource sharing networks, and its effects on
user participation and utilities. The evaluation showed interesting dependencies
between the effects of the scheme, the user type, and the user type distribution. The
second case study (Section 4.3) introduced the concept of co-operative infrastructures. It studied several contribution schemes for the provision of infrastructure
resources needed to run the platform by the users themselves.
Considering research question 1.2, these exemplary case studies show that the use
of a simulation-based approach to evaluate certain aspects of a sharing system,
such as an incentive scheme, can provide useful additional information about potential effects on the system. The first case study exemplified that dynamic effects
can be identified through the simulation tool, e.g., the effects of introducing an
additional trading constraint on different user types. With such an approach, potential changes to a Social Cloud can be evaluated before their implementation, and
can yield useful information about the predicted consequences. Additionally, the
second case study showed that entirely new scenarios can also be studied through
the simulation tool. For example, if a co-operative infrastructure should be introduced for a Social Cloud, the simulation approach can help to determine the most
effective contribution scheme to ensure a certain level of system availability.
Addressing research question 1.3, the second case study showed that the decision
for a certain contribution scheme depends on the user type distribution of the respective Social Cloud. If resource provisioning to the Social Cloud does not incur
considerable additional costs and the Social Cloud does not mainly consist of selfish users unwilling to participate in the co-op, a voluntary contribution scheme is
feasible. In contrast, if users are less willing to provide resources on their own,
an enforced contribution scheme can be necessary to ensure the applicability of a
co-op approach to host the platform infrastructure. Hence, in practical scenarios,
the identification of user type distributions is important to select the appropriate

126

Designing Incentive Schemes and Co-operative Infrastructures

contribution scheme. In addition, the importance of non-monetary incentives for
participation in a Social Cloud (in particular general helpfulness and reciprocity
aspects, see Section 3.3) indicates the general feasibility of such a co-operative infrastructure approach.
It is important to emphasize that such an approach is complimentary to other
methodologies such as prototyping or live experiments. Ideally, these methodologies complement each other and yield valuable insights which together lead to
a more holistic design process. To augment the case studies shown in this chapter, two prototypical implementations, a Social Storage Cloud (Chard et al., 2012)
as well as a Social Compute Cloud for sharing virtualized resources (Caton et al.,
2014), have already been implemented. These prototypes are particularly helpful in observing actual user acceptance, behavior and feedback with respect to the
general idea, system usability as well as specific implementation details. For example, user interaction can be observed and analyzed with respect to demographics
and other user characteristics. Complementary to the simulation and prototype
approaches, lab experiments can capture user behavior in certain market scenarios
and determine the motivation of potential users to contribute and share resources
in different application scenarios. This can be accompanied by surveys and questionnaires which are specifically suited to study the motivation of users, their incentives as well as determine the reason of their observed behavior.
Overall, Chapters 3 and 4 showed the relevance of non-monetary incentives for Social Cloud settings, and how the identified relevant incentives can be implemented
in incentive and contribution schemes that encourage user participation. A webbased survey showed that non-monetary motivations and incentives such as helpfulness and reciprocity considerations are more important for users than monetary
compensation. Additionally, the survey revealed that the setting of a Social Cloud
(e.g., sharing in a private compared to a professional setting) influences the relative
importance of certain incentives.
The findings are also relevant for the design of allocation mechanisms. Once users
of a sharing platform offer resources while others request these resources, the question arises how an allocation can be derived in this setting. Due to the low relevance of monetary incentives, the use of non-monetary allocation mechanisms is a
promising approach to retain the benefits of a centralized market allocation while

4.4. SUMMARY

127

focusing on preferences rather than (monetary) valuations for resources. This is the
focus of the next part of the thesis. Chapters 5 and 6, thus, take a market view of
the system and consider resource sharing where users have preferences with whom
they want to share. Specifically, several allocation mechanisms, their relative performance, as well as strategic implications will be the focus of these chapters.

Part III.
Two-Sided Matching in Social Clouds

Chapter 5.
Resource Allocation in Social Clouds
“Two properties of key importance for market design are stability, which encourages
groups to voluntarily participate in the market, and incentive compatibility, which
discourages strategic manipulation of the market.”
(Nobel Prize Committee, 2012)

I

N social settings, the exchange of resources is often driven by non-monetary
factors. This can be observed empirically by considering the various platforms
which use non-monetary incentives and mechanisms (such as trophies or reciprocity, see Section 3.1). The results of Chapter 3 also indicated that these types of
incentives are considered most important for users participating in a Social Cloud.
As identified in Section 2.3.2, two-sided matching provides the means to allocate
resources with a market-based approach that allows for the optimization of desirable criteria, such as fairness of welfare of an allocation, while at the same time does
not involve monetary transactions. Instead of (monetary) valuations for resources,
participating users specify a preference ranking that reflects their willingness to
share and exchange resources with other users. This chapter, therefore, considers
the application of preference-based matching algorithms in the context of Social
Clouds.
To introduce the relevant concepts, Section 5.1 provides the definitions and notation needed to study preference-based resource allocation. In particular, different
types of preference structures and performance metrics as well as their impact on
problem complexity are discussed. Section 5.2 provides an overview of the twosided matching literature and discusses several aspects that are commonly studied
in preference-based matching.

131

132

Resource Allocation in Social Clouds

Using this terminology, Section 5.3 describes existing algorithms for preferencebased matching. Most of these algorithms are specialized for specific scenarios,
i.e., focus on a certain combination of preference structures. While there are efficient (approximation) algorithms for certain scenarios, there are cases for which no
approximation algorithm exists.1 Additionally, only certain standard combinations
of optimization metrics are considered in the literature. However, there are cases in
which the flexible combination of optimization metrics might be useful. For example, in a Social Cloud the desired goals of an allocation might change over time. In
the beginning, the platform might focus on the benefits for the individuals to provide incentives for participation, whereas in later stages aspects such as fairness
with respect to different users might be of higher importance. Existing algorithms
are not able to provide such a flexibility and can only focus on a small combination
of metrics. Heuristic algorithms, on the other hand, allow for a flexible optimization of metrics and are suitable for different preference structures. Therefore, the
focus of this part of the thesis is the development and evaluation of heuristics for
two-sided matching. In particular, Genetic Algorithms (GA) (Goldberg, 1989) and
Threshold Accepting Algorithms (TA) (Dueck and Scheuer, 1990) are considered
for this case.
The sole existence of heuristics does not guarantee their usefulness, especially considering the quality of allocations calculated by the heuristics. Therefore, it is necessary to compare the performance of these heuristics with existing algorithms,
which is the goal of this chapter. Thus, the performance of the currently bestknown algorithms is compared to the heuristics in Section 5.4, which considers
several scenarios with different types of preferences. With the corresponding evaluation, research question 2.1 can be addressed:
R ESEARCH Q UESTION 2.2 ≺ P ERFORMANCE OF H EURISTICS  What is the per for mance of heuristics for preference-based matching compared to existing matching
mechanisms?
The contribution of this chapter is twofold. On the one hand, GA and TA heuristics
are defined for general two-sided matching scenarios and different optimization
goals. On the other hand, the chapter provides a performance comparison of these
1 In

fact, for some scenarios non-trivial approximation guarantees cannot exist (Manlove et al.,
2002; Halldórsson et al., 2003), which means that no algorithm can guarantee to always find
better than worst-case lower bounds on approximation.

5.1. PREFERENCE-BASED RESOURCE MATCHING

133

flexible heuristics compared to existing algorithms as well as the study of these
algorithms in various settings. For example, the performance of the algorithms in
the case of correlated preferences is a relevant problem that has not been examined
extensively before.

5.1. Preference-based Resource Matching
This section formalizes two-sided matching problems. Section 5.1.1 starts by introducing notation, definitions, and fundamental theorems on preference-based
matching, following standard literature. Building on this, Section 5.1.2 presents
the commonly considered performance metrics used to determine the quality of a
solution. Finally, Section 5.1.3 categorizes different problem types according to the
structure of the preference rankings, and discusses the computational complexity
of these problem types.

5.1.1. Definitions and Relevant Theorems
A two-sided matching model considers scenarios where users i are participants in
a market and want to share and exchange resources. It is assumed that a user i
cannot concurrently supply and demand the same resource type r. Therefore, it is
possible to split the users into the set of n X requesting users, Xr , and nY providing
users, Yr . Note that in this model matching is considered only within the same
resource type, i.e., the index r is omitted from all subsequent formulas. In total,
there are n X + nY users participating in the market. For easier notation, requesters
will denote users of side X and providers of side Y. Without loss of generality, the
index j will denote users of the opposite side.



Each user i has a preference profile Pi = Pi,j1 , . . . , Pi,jn over users j of the other
market side with whom they want to share resources, where Pi,j denotes the (ordinal) preference rank that user i has towards user j. The preference towards ∅ indicates the preference for being unmatched. Preference profiles are transitive and
can be asymmetric. The preference profiles represent transitive priority structures
= (i ) where each user of the opposite side is ranked according to its priority.
The asymmetric part i indicates a strict priority, whereas the symmetric part indicates an indifference. All users j with j i ∅ are said to be acceptable for user i

134

Resource Allocation in Social Clouds

(and vice versa). A common representation of preferences is through ranked order
lists. For example, if user i has preferences j2  j1 ∼ j3  j4 , the most preferred
alternative is j2 , which is preferred to both j1 and j3 , which are in turn preferred
to j4 . The corresponding preference ranks in this case would be Pi,j1 = 2, Pi,j2 = 1,
Pi,j3 = 2, and Pi,j4 = 4.2
A preference profile of user i is said to be complete if j  ∅ for all users j. If ∅  j
for some users j, the preference profile is said to be incomplete. This indicates that
user i prefers to remain unmatched rather than be matched to user j. A preference
profile is strict if for all users j, k of the opposite side j i k is asymmetric. If j ∼ k
for some users j and k, then the preference profile is said to have indifferences, or
ties.
Given the representation of users’ preferences and the supply and demand in the
market, the goal of two-sided matching is to find a match μ =
X, Y  that defines
which users are matched.
X, Y  consists of pairs
x, y with x ∈ X and y ∈ Y. In this
work, only one-to-one matches are considered, i.e., one requesting user is matched
to one providing user, and vice versa.3
A matching is a function μ : X → {Y ∪ ∅} that allocates one requesting user
from X (or ∅) to one providing user from Y (or ∅) and fulfills the constraints
∀i ∈ X : μ(i ) ∈ {Y ∪ ∅}, ∀i ∈ Y : μ−1(i ) ∈ { X ∪ ∅}, and ∀i, j = ∅ with μ(i ) = j :
¬∃k such that μ(k) = j or μ−1 (k) = i. In other words, each user is either matched
to another user or remains unmatched, and users can only be part of one matching
(i.e., no user can be matched to more than one other user). A mechanism implements μ for the given preference profiles.
The quality of a match μ is characterized by certain criteria. Fundamental to the
theory of two-sided matching is the question of whether users have the incentive to
deviate from a given match. If no two users have the bilateral incentive to deviate,
a match is called stable. For the definition of stability, the definition of blocking
pairs is necessary:
Definition 6 ( Blocking Pair ). A blocking pair is defined as a pair
x, y, x ∈ X, y ∈ Y,
such that
1. μ( x ) = ∅ or y  x μ( x ), AND
alternative ranking would be Pi,j4 = 3, in case only strict priorities increase the ranking.
discussion of many-to-one and many-to-many markets is provided in Section 5.2.

2 An
3A

5.1. PREFERENCE-BASED RESOURCE MATCHING

135

2. μ−1 (y) = ∅ is single or x y μ−1 (y), AND
3. x and y are mutually acceptable.
Blocking pairs are essential in characterizing the notion of stability in two-sided
matching. Throughout the chapter, following definition for stability will be used:
Definition 7 ( Stability ). A match is said to be stable if it contains no blocking pairs.
Stability infers that no user can find another user of the opposite side who prefers
it to its current partner, and both mutually benefit from the change of matched
partners. This is an important concept for theoretical and empirical reasons, and
the implications are discussed in the next section.
It is important to note that in case of preference profiles with indifferences, there are
other concepts of stability that can be used. Irving (1994) introduced the concepts
of strong stability and super stability. Both build on a slightly adjusted notion of a
blocking pair. In the case of weak stability (as used in this thesis), both partners
in the blocking pair have to strictly prefer each other to their current partner. For
strong stability, only one of the users has to strictly prefer the other user while the
second user can be indifferent between the options. In case of super stability, both
can be indifferent between their current match and the potential new match. It was
shown that while there always exists a weakly stable solution, this need not be the
case for strongly or super stable solutions (Irving, 1994). As weak stability is the
most often used in the literature, and strongly stable and super stable solutions do
not always exist (see Irving (1994)), this chapter will concentrate on this notion of
stability.
Considering the existence of solutions to the two-sided matching problem, the following result provides an optimistic outlook:
Theorem 1 ( Gale and Shapley, 1962, and Irving, 1994 ). For each two-sided matching
problem, potentially with incomplete preference lists and/or indifferences, there exists at
least one stable solution.
As mentioned before, stability is usually considered to be the most important characteristic of a match. Therefore, the existence of at least one stable solution to the
matching problem as shown in Theorem 1 is a very important and helpful result.

136

Resource Allocation in Social Clouds

On the other hand, the next definition shows that the number of different stable
solutions can be very large:
Theorem 2 ( Irving and Leather, 1986; Gusfield and Irving, 1989; Knuth, 1997 ).
Depending on the preference profiles, the number of different stable matches can be exponential.
Theorem 2 is an important result as it implies two aspects: 1) the number of different stable solutions can be very large, and an enumeration of these solutions might
not be feasible; 2) different stable solutions can be further evaluated considering
other performance criteria as well. Therefore, the next section introduces the commonly considered criteria which are used alongside stability to measure the quality
of a solution.

5.1.2. Performance Metrics for Two-Sided Matching
In standard two-sided matching scenarios, stability is often seen as the most important property. Yet, as seen in the last section the number of stable solutions for
a given set of preference profiles can be large, sometimes even exponential in the
number of users. For this reason, several other criteria are commonly used in conjunction with stability. For the matching problems considered in this chapter, the
following economic performance criteria are considered:4

Stability Stability is measured by the number of blocking pairs in a solution (see
Definition 6).

Welfare As a measure of the total “satisfaction” of users with respect to their
preferences, welfare is defined as the average rank of the matched partner for each
user.5 In formal terms:
Welfare =
4 The

∑i,j∈
X,Y  Pi,j + Pj,i
n X + nY

(5.1)

definitions of welfare and fairness scores are adapted from Gusfield (1987) and Iwama et al.
(2010). In addition to the presented metrics, regret is sometimes considered as well. It is defined
as the lowest preference rank that any user is matched with, and is a measure for how good a
solution is for the user with the lowest-ranked matched partner.
5 Welfare is also sometimes referred to as the most “egalitarian” solution.

5.1. PREFERENCE-BASED RESOURCE MATCHING

137

Note that lower numbers indicate better solutions, as the most preferred alternative
has rank 1.

Fairness Considering the welfare distribution of the two sides, fairness is measured as the difference of the average ranks of the matched partner. A higher fairness score reflects that users of one side are, on average, matched to partners with
a better rank than users of the other side, whereas scores around 0 reflect a more
equal welfare distribution. Formally:


 ∑i,j∈
X,Y  Pi,j ∑i,j∈
X,Y  Pj,i 


−
Fairness = 

nX
nY

(5.2)

Number of Matched Pairs For problems with complete preferences, the algorithms considered in this thesis always yield the maximum number of matched
pairs. In contrast, this property is lost by introducing incomplete preferences.
In such cases, the number of matched pairs is used as a quality metric for the
matches:
(5.3)
NumPairs = ∑ {
x, y | x = ∅ ∧ y = ∅ }

X,Y 

In case of incomplete preferences, finding the stable match with the highest number
of matched pairs is the most commonly considered combination of metrics. Even
though finding the welfare- and fairness-best stable solution can also be goals in
this case, the application of the welfare and fairness metrics to the incomplete preferences case requires the specification how unmatched users are handled for the
calculation of metrics. There is no standard in the literature how this case is handled, and most approximation algorithms focus on the mentioned combination of
stability and matched pairs. In the subsequent evaluation, the welfare and fairness
metrics will consider the matched users only.
For the given definitions, ranks are equally weighted, i.e., weights for different
preference rankings are not considered. This follows standard literature on twosided matching, yet cannot capture the fact that preference rank differences might
not be equidistant in real settings. For example, a user might consider the difference between the first and second rank differently than between the tenth and
eleventh. Although the subsequently described heuristics are easily adapted to a

138

Resource Allocation in Social Clouds

weighted preference setting, the unweighted setting is evaluated as algorithms for
the weighted settings do not exist for each relevant scenario.

5.1.3. Preference Structures and Computational Complexity
The goal of two-sided matching is to find an allocation that optimizes a certain combination of the aforementioned performance criteria. The complexity of finding an
optimal solutions depends on two input parameters: 1) the type of user preference
profiles, and 2) the combination of metrics that are used in the optimization.
To classify the preference types considered in a given problem, the case of complete
and strict preferences is abbreviated as SM (stable matching), complete preferences
with indifferences as SMT (stable matching with ties), and incomplete preferences
with indifferences as SMTI (stable matching with ties and incompleteness).
Theorem 3 ( Iwama et al. (1999); Manlove et al. (2002); Halldórsson et al. (2003) ).
If preferences are complete and include indifferences, finding a welfare-optimal or fairnessoptimal stable match is NP-hard, and also hard to approximate. If preferences are incomplete and include indifferences, finding a maximum size stable match is NP-hard, and finding a welfare-optimal maximum size stable match is NP-hard and hard to approximate.

There are two noteworthy implications when preference structures with indifferences (SMT) or incompleteness (SMTI) are considered. First, Theorem 3 shows
that finding an approximation algorithm for the welfare- of fairness-optimal stable solution is a hard problem itself, meaning that for these problems there is no
approximation algorithm that is able to guarantee a non-trivial solution quality.
This result exemplifies the importance of studying heuristics for finding solutions
in these cases. Second, if only indifferences are allowed in preference structures,
yet the preferences themselves are complete, then the adapted algorithms as well
as the heuristics guarantee that the maximum number of matched pairs is found.
This simply follows the fact that in case of completeness, every user of the other
side is acceptable, leading to a matching with maximum size (which is the size
of the smaller market side). As soon as incompleteness is introduced, the algorithms are not able to guarantee a maximum size matching anymore. In this case,
finding the maximum size stable matching is usually considered the most relevant

5.1. PREFERENCE-BASED RESOURCE MATCHING

139

objective, and finding the welfare- or fairness-optimal stable matching out of the
maximum size matching is considered a subordinate goal which further increases
problem complexity.
Table 5.1.: Computational Complexity of Two-Sided Matching Problems. SM indicates
complete and strict preferences, SMT complete with indifferences, and SMTI incomplete
with indifferences. Objectives are Stability (S), Welfare (W), Fairness (F), Number of
Matched Pairs (NM), or Multiple Objectives (MO).
Scenario Objective

SM

S
S&W
S&F
S&F
S&F
MO
S, W, Equity

SMT

S
S&W
S&F
S & W, S & F

SMTI

S
S & NM
S & NM
S & NM
S & NM
S & NM
MO

Complexity


O
n2 
O n4
NP-hard
NP-hard
NP-hard
NP-hard
NP-hard


O n2
NP-hard
NP-hard
NP-hard


O n2
NP-hard
NP-hard
NP-hard
NP-hard
NP-hard
NP-hard

Authors

Type

Abbreviation

Gale and Shapley (1962)
Irving et al. (1987)
Romero-Medina (2001)
Iwama et al. (2010)
Nakamura et al. (1995)
Vien and Chung (2006)
Kimbrough and Kuo (2010)

Exact
Exact
Exact
Approx
Heuristic
Heuristic
Heuristic

DA
WO
FE
-

Gale and Shapley (1962)
Irving et al. (1987)
Halldórsson et al. (2003)
Haas et al. (2013)

Exact
Exact
Approx
Heuristic

DA
WO
FE
GA

Gale and Shapley (1962)
Halldórsson et al. (2007)
McDermid (2009)
Paluch (2012)
Király (2011)
Gelain et al. (2013)
This Work

Exact
Approx
Approx
Approx
Approx
Heuristic
Heuristic

DA
Shift
McDermid
GSModified
Király
LocalSearch
GA, TA

Table 5.1 provides an overview of the computational complexity of the matching
problems relevant for this work. For the baseline case of complete and strict preferences (SM), it is possible to find a stable match, as well as the stable match with
the best welfare, in polynomial time. Yet even with these rather strict assumptions about preference types, finding the stable match with the best fairness score
is NP-hard. Once either the completeness or the strictness assumption is relaxed,
the complexity of finding stable solutions that optimize an additional performance
criterion increases. In case of complete preferences with indifferences (SMT), finding the welfare-optimal solution becomes NP-hard. The algorithm of Irving et al.
(1987) can still be used to calculate a solution, yet as the solution quality depends
on the way that the indifferences are broken, it cannot guarantee an optimal solution anymore.6 Computing the fairness-optimal solution remains NP-hard. In
6 To

guarantee optimality, the algorithm would have to be applied to all possible ways to break
the indifferences. In the general case, this number can be exponential in the lengths of ties and
number of users.

140

Resource Allocation in Social Clouds

the general case of incomplete preferences with indifferences (SMTI), the standard
goal is to find a stable solution of maximum size. This problem is also NP-hard,
and several approximation algorithms have been developed for this case.
Considering preference structures, besides the classification into SM, SMT and
SMTI, preferences can be categorized according to their correlation, i.e., the degree to which the preferences of multiple users are correlated. On the one hand,
preferences are uncorrelated if each user’s preferences are independent of other
users’ preferences, e.g., if preferences are created randomly. On the other hand, Celik and Knoblauch (2007) propose a method to create correlated preferences. Given
a certain threshold percentage x, the first x preferences are common among all participants of a side, and the last n − x preferences are common as well (the actual
order can be different, of course). This introduces a correlation in the set of preferences which can have an effect on the algorithm performance.7 In the subsequent
evaluation, uncorrelated preferences will be used as a baseline, and preference correlation according to the mentioned method will be applied as well to study its
effects on the matching outcome.

5.2. Related Work
The literature on two-sided matching, beginning with the seminal paper of Gale
and Shapley (1962), has emphasized stability as the design objective for two-sided
matching. Under the assumptions that preference rankings are strict and independent (of preferences of other individuals),8 Gale and Shapley (1962) introduced two
of the standard problems in two-sided matching, the College Admissions problem
and the Marriage Market. They also provided the first description of the DeferredAcceptance (DA) algorithm, which under the mentioned assumptions is able to
find at least one and at most two stable matches rapidly (in polynomial time).9
The literature on two-sided matching has grown considerably since Gale and Shapley (1962), as have the applications of two-sided matching. Roth (2008) and Abdulkadiroglu and Sönmez (2010) provide extensive overviews on the field of two7 Another

possibility is to study intercorrelated preferences (see Boudreau and Knoblauch (2010)),
however this is left for future work.
8 In particular, no “two-body problems” in which two or more individuals preferentially interact,
such as in spouses preferring to locate near each other, were permitted.
9 As Roth (2008) notes, certain labor markets had been managed successfully since the 1950’s using
a version of DA, but Gale and Shapley were unaware of this at the time.

5.2. RELATED WORK

141

sided matching and related concepts such as one-sided matching. Broadly speaking, there have been several main areas of research:
1. Types of Matching Problems
2. Preference Structures
3. Alternative Design Objectives
4. Computational Complexity.

Types of Matching Problems As mentioned earlier, the focus of this work is 1 : 1
resource allocation, i.e., one user of side X is matched to one user of side Y, and vice
versa. This model, often termed a “Marriage Market”, was first defined by Gale
and Shapley (1962) and has received considerable attention. There are, however,
other types of models that are noteworthy to discuss. For example, if providing
users of side Y have the capacity to serve (or to be matched with) several users of
side X, the problem becomes a 1 : n matching. The standard model in this case is the
College Admission Model (see e.g. Gale and Shapley (1962); Roth (1985); Balinski
and Sönmez (1999)), where users of side X request one unit of a resource (e.g., a
college space), and users of side Y have a certain capacity, cy ≥ 1, of resources they
can provide. It is noteworthy that some of the algorithms used for 1 : 1 matching
problems can be used for 1 : n problems through certain adjustments as well, and
that several of the important theorems and implications presented in the previous
chapters also hold in this case. For the more general model of m : n matching, Roth
and Sotomayor (1992) provide a good overview of algorithms that can be used in
this case.

Preference Structures Regarding preferences, the DA and much of the subsequent literature focuses on problems with strict preference orderings. If ties (indifferences) are introduced into the problem, certain characteristics of the algorithms
can no longer be guaranteed. For example, in order to use the standard algorithms
such as DA, the ties have to be broken first as the DA only allows strict preferences
as input. Erdil and Ergin (2006, 2008) introduced an algorithm that can cope with
ties in preferences. Their algorithm tries to find potential Pareto-improvement cycles in a given solution which might improve the overall quality of the solution. As
Gusfield and Irving (1989, p.219) note, however, many of the strong results for DA

142

Resource Allocation in Social Clouds

and related algorithms depend upon strict preference orderings, and characterizing stable matches under partial ordering remains a largely open problem. Scott
(2005) considers several variants of the standard, symmetric matching problem,
specifically concentrating on different concepts of stability. Abdulkadiroğlu et al.
(2009) study the effect of tie-breaking on the efficiency of the DA mechanism, and
also consider the stability-cost of finding an efficient matching which does not have
to be stable. They show that in some cases the welfare of participants of one side
can be improved, which leads to the introduction of a potentially large number of
blocking pairs.
Considering the type of preferences, the previously introduced definitions follow
standard literature and specify a qualitative preference ranking, i.e., for every two
alternatives the ranking determines whether the first alternative is better, worse,
or equal to the second alternative. They do not, however, allow for a quantitative
comparison of how much better or worse an alternative is. For example, from the
ranking a1 > a2 > a3 it cannot be inferred how much better a1 is compared to a3 .
The concept of weighted preferences extends the previous qualitative definitions and
assigns a weight (or score) to each preference rank. Such a weighting of preferences
has been proposed by Irving et al. (1987). They present adapted algorithms for certain matching problems, using the standard performance metrics. However, Pini
et al. (2011b) argue that the performance metrics themselves have to be adapted to
capture preference weights. Overall, as algorithms for weighted preferences have
only been developed for certain preference structures and a quantitative weighting
of preferences might not always be necessary, this work focuses on the standard
definition of qualitative preferences in two-sided matching.

Objective Functions Regarding design objectives for matching, other than stability, it was shown early on that DA is heavily biased as it finds the optimal stable
match for one side, and the pessimal stable match for the other side (Roth and Sotomayor, 1992). This raises the question of finding stable matches (or matches with
only a few unstable pairs of pairs) for the sake of other criteria, such as fairness
and social welfare. For strict and complete preferences, Irving et al. (1987) efficiently compute the welfare-optimal stable match. Axtell and Kimbrough (2008)
discuss trade-offs between stability and welfare, and Klaus and Klijn (2006b) study
(procedural) fairness and stability. Iwama et al. (2010) propose an algorithm that
approximately yields the fairness-optimal stable matching. Using heuristics such

5.2. RELATED WORK

143

as Genetic Algorithms, Kimbrough and Kuo (2010) show that they can yield superior solutions for welfare and fairness if a certain instability is allowed. Other
approaches that look at different or multiple objectives for certain matching problems include Vien and Chung (2006), Klaus and Klijn (2006a), Pais (2008), Pini et al.
(2011a), and Boudreau (2011) which also consider economic criteria such as welfare. From the perspective of what metrics are most relevant in practice, Echenique
and Yariv (2011) study one-to-one matching in experimental settings where participants have full information about the preferences. They find that, among other
things, stable matches are the most prevalent outcome, and that the cardinal representation of the ordinal preferences impacts which of the different matches is
selected.

Computational Complexity The issue of computational complexity arises once
preferences with incompleteness or indifferences are considered. On one hand,
the number of stable matches can be large, sometimes exponential in the size of
the problem (Gusfield and Irving, 1989; Knuth, 1997), and it has been shown that
the two-sided matching problem in general is #P-complete (Gusfield and Irving,
1989, p.157). Furthermore, for the considered SMT and SMTI preference structures,
the respective optimization problem is NP-hard (Halldórsson et al., 2007), and for
certain scenarios such as finding the maximum number of matched users Integer
Program formulations exist (Iwama et al., 2010). As discussed in Section 5.1.3, for
strict and complete preferences there are polynomial-time algorithms to compute
the welfare-optimal (Irving et al., 1987) and approximately fairness-optimal solutions (Iwama et al., 2010). However, by introducing indifferences and/or incompleteness, the problem of finding the welfare-optimal, minimum-regret or fairnessoptimal stable match becomes NP-hard, and sometimes even hard to approximate
(Halldórsson et al., 2007). Due to this complexity, heuristics have been studied to
obtain solutions to the matching problem, the GA being a prominent example. For
example, Nakamura et al. (1995) study whether a GA can yield stable matches with
higher fairness than the DA solutions, yet do not consider indifferences or other objectives. Aldershof and Carducci (1999) describe a GA to compute stable solutions
from random initial assignments, with stability as the sole objective. Furthermore,
both Kimbrough and Kuo (2010) and Vien and Chung (2006) compare a GA with
multiple objectives to the standard algorithms, yet neither of them consider indifferences in preferences.

144

Resource Allocation in Social Clouds

Summary
The literature review shows that while two-sided matching is a well-studied problem in the literature, there are still some aspects that need further investigation.
Table 5.2 provides an overview of the most relevant related work for 1 : 1 resource
alloaction. Especially considering the flexibility towards optimizing different or
changing objectives, existing algorithms mostly focus on a given set of optimization goals, and for certain combinations of goals no efficient optimization or approximation algorithms exist. As the calculation of the optimal solution is, in general, NP-hard, heuristics are particularly well suited for scenarios with multiple or
flexible objective functions and different preference settings as they can be adapted
to cover the respective scenario. Previously developed heuristics, however, have
either focused on the setting of strict preferences or a certain combination of optimization goals. Hence, the heuristics developed in this thesis bridge this research
gap by providing the flexibility to cover different preference structures.
Table 5.2.: Overview of Related Work
Authors

Flexible Preference
Structures

Flexible Objective
Function

Computational
Complexity

Iwama et al. (2010)
Gale and Shapley (1962)
Irving et al. (1987)
Iwama et al. (2010)
McDermid (2009); Király
(2011); Paluch (2012)
Erdil and Ergin (2008)
Vien and Chung (2006)
Kimbrough and Kuo (2010)
Goal

5.3. Algorithms for Preference-based Matching
Most algorithms found in the literature concentrate on finding stable matches under certain conditions. Depending on the quality guarantees that an algorithm
offers, they can be distinguished into exact, approximate and heuristic algorithms.
Exact algorithms yield the optimal solution for a given scenario, and approximate

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

145

algorithms guarantee that the solution quality is within a certain bound of the optimal solution. Heuristic algorithms, in general, do not provide such a quality
bound, yet have other advantages such as the flexibility to consider multiple simultaneous objectives. This section introduces the algorithms considered for this
chapter. Section 5.3.1 starts with the description of exact algorithms and the scenarios they can be applied in, and Section 5.3.2 continues with an overview of state-ofthe-art approximation algorithms. Finally, Section 5.3.3 introduces and describes
the heuristic algorithms that are focus of this chapter.

5.3.1. Exact Algorithms
As seen in Section 5.1.3, for certain preference structures the calculation of the optimal solution is possible in polynomial time. For strict preferences the Deferred
Acceptance (DA) algorithm by Gale and Shapley (1962) can be used which always
yields a stable outcome. Additionally, in this case the welfare-optimal (WO) algorithm by Irving et al. (1987) yields the welfare-optimal (or most egalitarian) stable
solution in polynomial time. In the case of indifferences, a tie-breaking rule has to
be applied first in order to apply these algorithms. The tie-breaking rule greatly affects the quality of the resulting matching and, in general, applying the algorithms
after tie-breaking does not guarantee a good solution. This section describes the
respective algorithms. In addition, the matching problem is formulated as a Linear
Program for the SM case and as Integer Program for the SMTI case.

Formulation as Optimization Model The two-sided matching problem can be
formulated as an optimization problem. For the standard model of strict preferences, Vande Vate (1989), Rothblum (1992), and Roth et al. (1993) introduced a linear programming (LP) formulation to obtain stable matches, which consists of a set
of linear constraints:

146

Resource Allocation in Social Clouds

max( x,y)

∑

i ∈ X,j∈Y

∑ zi,y ≤ 1

∀y ∈ Y

(5.5)

∑ zx,j ≤ 1

∀x ∈ X

(5.6)

∀
x, y ∈ A

(5.7)

z x,y ≥ 0

∀ x ∈ X, y ∈ Y

(5.8)

z x,y = 0

∀ ( x, y) ∈ ( X × Y ) \ A

(5.9)

i∈X
j ∈Y

∑

j x y

z x,j +

∑

i y x

(5.4)

zi,j

zi,y + z x,y ≥ 1

In their model, the decision variable zi,j determines whether users i and j are
matched (zi,j = 1) or not (zi,j = 0), and A defines the set of acceptable matched
pairs. The optimization function (5.4) maximizes the number of matched pairs.
Equations (5.5) and (5.6) ensure that each user is only allocated once. Equation
(5.9) prohibits matched pairs which are not mutually acceptable, and Equation (5.7)
guarantees that the solution is stable. Although Equation (5.8) is formulated as a
linear constraint, the properties of the optimization problem ensure that a solution that satisfies constraints (5.5) - (5.9) is in fact an integer solution of the given
optimization problem (Rothblum, 1992; Roth et al., 1993).
For the case of indifferences in preferences, the previous formulation has to be
adapted. In particular, applying the formulation to a problem with indifferences
might not yield valid integer solutions anymore, which makes an adaptation of
constraint (5.8) necessary. This also implies that the corresponding optimization
problem is an Integer Program (IP), which are, in general, NP-hard to solve.10 Using a formulation similar to Iwama et al. (2014)11 , the SMTI one-to-one matching
problem can be formulated as follows:

10 This

is in accordance with the previous theoretical findings in Table 5.1.
difference in this formulation is in Equation (5.13), which stems from a slightly different
definition of the  relation. Conceptually, however, the two formulations are equivalent.

11 The

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

max( x,y)

∑

i ∈ X,j∈Y

∑ zi,y ≤ 1

∀y ∈ Y

(5.11)

∑ zx,j ≤ 1

∀x ∈ X

(5.12)

∀
x, y ∈ A

(5.13)

∀ x ∈ X, y ∈ Y

(5.14)

∀ ( x, y) ∈ ( X × Y ) \ A

(5.15)

j ∈Y

∑

z x,j +

∑

i y x

(5.10)

zi,j

i∈X

j x y

147

zi,y + z x,y ≥ 1
z x,y ∈ {0, 1}
z x,y = 0

As before, the optimization goal is to maximize the number of matched pairs
(Equation 5.10). Other optimization functions can be used as well, provided
they are linear. For example, for complete preferences with indifferences the
welfare-optimal solution can be obtained by using the optimization function



∑i∈ X,j∈Y Pi,j + Pj,i zi,j . Constraints (5.11), (5.12), and (5.15) are unchanged, and
constraint (5.13) ensures a (weakly) stable solution.

Deferred Acceptance Algorithm
Being introduced in the seminal article by Gale and Shapley (1962), the Deferred
Acceptance (DA) algorithm has been widely used and adapted for applications
in research and practice. It was first introduced for the simple Marriage Market
(symmetric matching instances with n X = nY ), and subsequently adapted for other
matching problems such as College Admission problems. The strengths of the DA
are its fast runtime, applicability for different preference structures, and simplicity.
Considering complexity, the DA runs in O(n2 ) where n is the size of a (symmetric) market side, which makes it the fastest deterministic algorithm among the described algorithms to solve the matching problem. It can be easily adapted to cope
with incomplete preferences, and indifferences are usually handled by breaking the
ties first before applying the DA in its standard formulation. The pseudocode of
the DA is given in Algorithm 1.

148

Resource Allocation in Social Clouds

Algorithm 1: Pseudocode of Deferred Acceptance Algorithm (Gale and Shapley, 1962)

1
2
3
4
5
6
7
8
9
10

11
12
13
14
15
16

Data: Preference Profiles
Result: Stable match μ
begin
If some preference profiles are not strict, break the ties to get a strict ranking;
while Unmatched users of side 1 who have not proposed to all acceptable users in their
preference list do
select a user of side 1 that is unmatched and has acceptable partners to which
they didn’t propose yet;
user proposed to partner that is highest on their list;
if partner is unmatched and finds user acceptable then
partner temporarily accepts proposal;
else if partner is matched and prefers user to current match then
partner temporarily accepts proposal;
formerly matched user becomes unmatched and removes partner from
preference list;
else
user removes partner from preference list;
end
end
All users are matched to their current partner
end

Despite these advantages, the DA has several limitations as well. Firstly, although
it always yields a stable solution, for a given set of strict preferences it can only
find at most two different stable solutions by switching the side that starts the algorithm. As there are up to exponentially many stable solutions (Knuth, 1997), the
solution quality of the DA with respect to other performance metrics such as welfare is not immediately clear. Secondly, the DA yields particularly unfair solutions
in the sense that it calculates the best stable solution for the starting side and the
pessimal stable solution for the second side. This has to be considered especially
when fairness aspects between the two market sides are of importance. Lastly, in
case of incomplete preferences the DA does not guarantee a matching of maximum
size. Albeit an undoubtedly useful algorithm, for these reasons other algorithms
have been developed over time to address these issues.

Welfare-Optimal Algorithm
For the problem of finding a welfare-optimal stable solution in case of strict and
complete preferences, Irving et al. (1987) present an algorithm that yields a solu-

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

149

tion in polynomial time (henceforth called Welfare-Optimal algorithm, WO). If n is
the size of one side in a symmetric setting, the runtime of their algorithm is O(n4 ).
The pseudocode of WO is shown in Algorithm 2. The algorithm uses the concept
of rotations, which were studied in the context of a symmetric 1:1 market by Irving and Leather (1986). They showed that all stable solutions of a given matching
problem can be identified by the use of these rotations. Essentially, the algorithm by
Irving et al. (1987) uses this rotation concept in combination with graph-theoretic
algorithms in order to find the sequence of rotations that yield the welfare-optimal
stable solution. Similar to the DA, WO can be used in case of strict preferences
with or without indifferences by breaking the ties before the start of the algorithm.
However, in case of indifferences WO only guarantees the welfare-optimal stable
solution for a given tie breaking; the global welfare-optimal matching cannot be
guaranteed.
Algorithm 2: Pseudocode of Welfare-Optimal Algorithm (Irving et al., 1987)

1
2
3
4
5
6
7

Data: Preference Profiles
Result: Welfare-optimal stable match μ
begin
Calculate shortlists for each user;
Calculate all rotations given the shortlists;
Create graph from weighted rotation poset P;
Compute maximum-weight closed subset;
Eliminate all rotations in maximum-weight closed subset;
end

The O(n4 ) algorithm by Irving and Leather (1986) was later improved to a
O(n2.5 log n) algorithm by Feder (1992). As the focus of the subsequent evaluation is the solution quality rather than the runtime (the quality is the same as both
are optimal algorithms), only WO is considered in the evaluation.

5.3.2. Approximation Algorithms
As it is not possible to calculate optimal solutions through polynomial-time algorithms for all matching problems, in certain cases approximation algorithms can be
used to calculate a solution with a specified quality bound. Examples of the problem types where an exact solution cannot be computed in polynomial-time is finding the fairness-optimal stable solution (even for strict and complete preferences),
and finding the maximum size stable matching in case of incomplete preferences

150

Resource Allocation in Social Clouds

(see the overview in Table 5.1). For these problems, approximation algorithms have
been developed.

Fairness Approximation
Finding the fairness-optimal stable solution is NP-hard even for strict and complete preferences. Hence, Iwama et al. (2010) propose an approximation algorithm
(henceforth called Fairness-Equal, FE) that yields stable solutions with a certain
quality bound on the fairness score. The FE algorithm works with a parameter 
that specifies the fairness bounds within which the near-optimal fair solution is intended to be found. Without loss of generality, the implementation of FE used for
the evaluation assumes  = 0.1. Clearly, the specification of this bound involves a
certain trade-off. On one hand, if the bound is too tight, it is possible that no solutions will be found within the bound. On the other hand, if the bound is too large,
the solution quality with respect to the fairness score might be worse, as solutions
are accepted which might be further away from a perfectly fair solution.

MaxPairs Approximation
As described earlier, in case of incomplete preferences it is NP-hard to find the
stable match of maximum size. Several approximation algorithms have been suggested for this case. As the trivial approximation ratio in this case is 2, i.e., the worst
case size of a solution is half the size of the optimal solution, any approximation
ratio smaller than 2 is desirable.
• Shift: Halldórsson et al. (2007) describe an approximation algorithm for this
case, in the following abbreviated as Shift. For certain preference structures, this algorithm provides non-trivial quality bounds for finding the stable
match of maximum size. Shift operates through breaking indifferences in a
systematic manner and applying the DA on the resulting set of strict preferences. In particular, if indifferences occur on both sides of the market, Shift
guarantees non-trivial quality bounds if the length of indifferences is at most
2.
• McDermid: McDermid (2009) presents an algorithm with a 3/2 approximation ratio, which is the best known approximation ratio for the general case
without restrictions on tie lengths.

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

151

• Király: Similarly, Király (2011) presents an algorithm with a 5/3 approximation ratio, which has a slightly better runtime than that of McDermid (2009).
• GSModified: Paluch (2012) presents another algorithm with the same runtime
and approximation ratio as the algorithm by Király (2011).
These algorithms are used in the subsequent comparison. If ties are only one-sided,
(Iwama et al., 2014) provide an algorithm that guarantees an approximation ratio of 25/17. Irving and Manlove (2008) present algorithms to approximate stable
marriage and hospital/residents problems. As the former puts a considerable restriction on the preferences, and the latter has a worse approximation ratio than
McDermid, Király and GSModified, they are not considered in the evaluation.

5.3.3. Heuristics
In addition to the mentioned algorithms, the use of heuristics to find solutions to
the matching problem is investigated. In this chapter, besides improvement cycles
as suggested in the literature (Erdil and Ergin, 2008), two different heuristics are
used: a Genetic Algorithm (GA) (Goldberg, 1989) as an example of a evolutionary (meta)heuristic, and a Threshold Accepting algorithm as an example of a local
search heuristic. In general, heuristics can be used to find (stable) solutions from
random initial assignments (see Aldershof and Carducci (1999); Kimbrough and
Kuo (2010)), or to improve an initial stable matching by trying to retain stability
and increasing other performance criteria (see Haas et al. (2013)).

Improvement Cycles
In the case of (potentially incomplete) preferences with indifferences, Erdil and
Ergin (2008) suggest an algorithm that computes improvement cycles for a given stable solution. This algorithm, called Requester-Optimal-Stable-Matching (RSMA)
throughout the chapter, looks for ordered sequences of matched pairs, such that
each requester (weakly) prefers to be matched to the provider of the next pair.
If such a sequence of matched pairs exists, switching each requester to the next
provider on the respective sequence element yields another stable match that is
Pareto-superior for the requesting users (though not necessarily for the providers).
In case the sequence of pairs contains a pair with an unmatched requester and
a pair with an unmatched provider, the sequence is called an improvement cycle, otherwise an improvement chain. It is easy to see that an improvement cycle

152

Resource Allocation in Social Clouds

increases the number of matched users, whereas an improvement chain yields a
match with the same number of matched users.
While RSMA guarantees that users of one side are not worse of from the improvement, this is not necessarily true for users of the second side. For such cases,
Erdil and Ergin (2006) suggest an alternative algorithm, the Efficient-and-StableMatching-Algorithm (ESMA), which focuses on improvements in which both sides
are (weakly) better off. As this restricts the potential chains and cycles that can
be found, the subsequent evaluation only considers RSMA, as initial results have
shown that applying ESMA does not significantly improve the solution in but a
small number of cases.

MaxPairs Heuristic
Gelain et al. (2013) present local search heuristics to solve the generalized stable
matching. They start with solving the relaxed version of the problem (assuming
complete preferences), thereby potentially introducing instability, and then deleting unstable pairs through an iterative process until stable solutions are found. Initial tests showed that its performance with respect to solution quality is worse than
the heuristics proposed in this chapter, which is why it has not been considered for
further evaluation.

Genetic Algorithms
Starting with the groundbreaking works of Goldberg (1989) and Holland (1992),
Genetic Algorithms (GA) have been widely used for and successfully applied to
many optimization problems. Specifically, they have been found to perform well
in case of large search spaces (Duffy, 2006; Kimbrough and Kuo, 2010; Haas et al.,
2013).
GAs start with a (usually randomly created) initial set of potential solutions (the
population) and evolve this solution set by applying certain mathematical operations on them. The quality of a potential solution is determined by its fitness, i.e.,
how well it performs with respect to the given objective function. The aim of this
type of evolutionary algorithm is to create new solutions which are better than the
initial population. Each potential solution, a chromosome, consists of several genes,
where each gene represents the value of an attribute of the solution. Traditionally, chromosomes and genes are encoded as bit-strings, although integer- or realvalued genes are also commonly used in practice. In order to find solutions with

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

153

high fitness values, the GA evolves the population by applying the mentioned operations, yielding a new population of potential solutions.12 Usually, there are three
main genetic operators used to obtain the new population, which are specified as
follows (see Goldberg (1989)):

Selection Through the selection operator, a certain number of chromosomes of
the old population are selected into an intermediate population, and the subsequent operations are performed only on this intermediate population. Intuitively,
this means that only a certain number of solutions are allowed to reproduce, and
usually only the fittest solutions are selected in this step. The selection itself can be
implemented in various ways. In practice, many GA implementations use either
a tournament or a roulette-selector. A tournament selector of size k randomly selects k solutions out of the population and selects the fittest of these k solutions. In
contrast, a roulette-selector assigns probabilities to individual solutions, where the
probability of being selected is proportional to the solution’s fitness.

Crossover The crossover operator was developed as an analogy to biological reproduction. Usually, crossover is only applied to a certain percentage of solutions
of the intermediate population. The crossover itself is performed in two steps.
First, two solutions of the intermediate population are selected. Second, certain
parts of the two selected solutions are swapped. There are several commonly used
crossover operators that determine which parts of the chromosomes are swapped.
As some crossover operators might create invalid solutions depending on the specific encoding of the chromosomes, the right crossover operator has to be chosen
for the given problem. The newly created solutions, called offspring, then usually
replace the parent solutions.

Mutation After crossover, the mutation operator randomly changes the values of
some genes according to a specified mutation probability. The rate of mutation is
essential for the performance of the genetic algorithm. If the mutation rate is too
12 An

important theoretical result on the efficiency of a GA is the schema-theorem, also referred to
as the building block hypothesis. The theorem states that schemas, defined as a subset of genes
which are inherent to fit chromosomes and that are found across several chromosomes, are exponentially more often propagated in successive generations (Goldberg, 1989). In other words,
inherent characteristics of successful solutions are exponentially more often found in next generations of the population. This is important for the matching problem, as it states that successful
matched pairs are propagated to the new solutions.

154

Resource Allocation in Social Clouds

low, the chance of getting stuck in local optima increases. In contrast, setting the
mutation rate too high does not allow the algorithm to settle and converge to a
solution due to the continuous adjustments through mutations. Hence, care must
be taken in selecting a mutation rate, as this fundamentally affects the outcome of
the genetic algorithm.
For the matching problem, each chromosome represents a solution
X, Y . A chromosome consists of several genes, where each gene encodes a provider-requester
match
x, y of the solution. In other words, when a solution has m matches, the
chromosome has m genes, and each gene consists of two identifiers, one for the
provider, one for the requester.13 As fitness functions for two-sided matching,
the maximization of stability, welfare, fairness, number of pairs, or a combination
thereof are used. As each chromosome only encodes a set of matched pairs, the
preferences of the users are needed to evaluate its fitness.
In order to improve the fitness of the solutions, the two described genetic operators are applied after the fitness evaluation in order to derive new, potentially
better-performing solutions. For crossover, the cycle crossover operator (Goldberg,
1989) creates new potential solutions by combining two parent solutions. This type
of crossover operator ensures valid solutions for the given encoding by exchanging
the same set of IDs for the two chromosomes.14 The mutation operator, given a certain mutation probability, depends on the type of preferences and optimization objective. In case of complete preferences with indifferences, it randomly selects two
genes (matched pairs) of a given chromosome and exchanges either the requester
or provider identifiers to create a new chromosome. In case of incomplete preferences, it randomly selects a gene with an unmatched requester, a randomly chosen amount of genes representing matched pairs, and a gene with an unmatched
provider. This is done for a similar reason as the RSMA algorithm by Erdil and
Ergin (2008): to potentially find an improvement cycle that, when each requester
is matched to the provider of the next pair in the cycle, increases the number of
matched pairs. This specification of crossover and mutation operators ensure valid
solutions which are not necessarily stable. As the subsequent evaluation focuses on
stable matching algorithms, blocking pairs are discounted in the objective function
that the maximum number of genes is n X + nY , which represents the case that each user is
unmatched.
14 The use of standard crossover operations might yield solutions where users are matched multiple
times. Hence, only the class of crossover operations that are applicable for enumeration settings
are of interest.
13 Note

5.3. ALGORITHMS FOR PREFERENCE-BASED MATCHING

155

of the GA, which ensures that newly created, unstable solutions are not likely to be
propagated through the evolution rounds. The population is evolved using these
operators over a given number of rounds. For the selection of solutions that form
the new population in the next evolution round, a tournament selector of size 2 is
used. Additionally, through an elitist selection the 5 best solutions are guaranteed to
be transferred to the next population. Algorithm 3 shows the pseudocode for the
GA.15
Algorithm 3: Pseudocode of Genetic Algorithm based on (Goldberg, 1989;
Holland, 1992)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Data: Preference Profiles
Data: Crossover and mutation probabilities
Data: Number of evolution rounds and size of population
Data: Number of best solutions retained: nbest
Data: Optional: Starting Solutions
Result: Match μ
begin
Create initial population: solution set ← initial population;
for i ← 1to number evolution rounds do
temporary solution set ← perform crossover operation;
for solution ∈ temporary solution set do
mutate solution according to mutation probability;
end
solution set ← ∅;
solution set ← nbest solutions;
while size of solution set < size of population do
run tournament selection and include winner in solution set;
end
end
μ ← best solution of population;
end

To show the actual encoding and how the genetic operators work, an example for
the encoding, the cycle crossover, and the mutation operator are provided in Appendix C.

Threshold Accepting Algorithms
The second class of heuristics studied in this thesis are Threshold Accepting (TA) al15 The

number of evolution rounds can be dynamically adjusted if the available computation time
is limited, e.g., in settings where a solution has to be calculated in short time. If the calculation
of a solution is not time-critical, the solution quality can be potentially increased by evolving the
population over more rounds.

156

Resource Allocation in Social Clouds

gorithms (Dueck and Scheuer, 1990). TAs are an example of local search heuristics,
where a given starting solution is improved by sequentially adjusting the solution
and accepting adjustments within a certain threshold. TAs are conceptually similar
to Simulated Annealing approaches and have been successfully applied in many
optimization problems. Depending on the definition of the solution adjustment
per step, TAs try to improve a given solution and hence are suitable for finding
(especially local) improvements. Compared to GAs which work on a population of
different solutions in the solution space, the performance of a TA depends on the
quality of the starting solution. However, it is more flexible on incrementally improving this solution than the GA which has to rely on mutations to find improvements once it settled on a local or global optimum. The subsequent evaluation
studies its applicability in the case of two-sided matching.
The general procedure is shown in Algorithm 4. Given a starting solution, a set
of thresholds is defined. For each of these thresholds, a certain number of adjustments are sequentially performed on the solution. For the matching problem, the
adjustment is similar to the mutation operator of the GA. For complete preferences,
it selects two matched pairs and switches either the requesting or providing users.
For incomplete preferences, it selects an unmatched user of side X, a randomly
drawn number of matched pairs, and an unmatched user of side Y (thus forming a cycle of matched pairs), and replaces the users of side X with the respective
user of the previous matched pair in the cycle. An adjustment is accepted as new
(temporary) solution if it does not decrease the solution quality by more than the
threshold (compared to the current solution). Thresholds are reduced over time
such that convergence to a (local) optimum becomes more likely, whereas the initial thresholds are set to avoid being stuck in a local optimum too soon.

Genetic Algorithm with Subsequent Threshold Accepting
GAs tend to sample (especially large) search spaces better than local search heuristics as they start with a potentially diverse set of solutions. However, as mentioned
before the incremental improvement of a given solution depends on mutations.
This might not be as efficient as using local search heuristics that aim at improving
a given solution. Hence, the combination of the two approaches is also studied,
where the GA is used to find a good starting solution for the TA, which then tries
to further improve this solution. For the purpose of evaluation, this combination
of heuristics, which represents a memetic algorithm, is abbreviated as GATA.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

157

Algorithm 4: Pseudocode of Threshold Accepting Algorithm based on (Dueck
and Scheuer, 1990)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

Data: Preference Profiles
Data: Number of thresholds
Data: Number of repetitions per thresholds
Data: Optional: Starting Solution μstart
Result: Match μ
begin
if μstart = ∅ then
μcurrent ← create starting solution;
else
μcurrent = μstart ;
end
for i ← 1to number of thresholds do
for j ← 1 to number of repetitions per threshold do
μ adj ← apply adjustment to μcurrent ;
if score μ adj ≤ μcurrent + current threshold then
μcurrent ← μ adj ;
end
end
end
μ ← μcurrent ;
end

5.4. Performance of Matching Algorithms
Algorithms such as GA and TA have an inherent flexibility to be adjustable to many
different optimization functions and combinations of metrics that need to be optimized. This stands in contrast to specialized algorithms, such as the approximation algorithms for calculating the solution with the maximum number of matched
pairs, which focus on one specific scenario or set of goals. The aim of this section,
therefore, is to provide a comparison of the performance of the different algorithms
in specific settings.
For the evaluation, a simulation-based approach is used. This provides the flexibility to test the performance of the algorithms in various different settings and
perform sensitivity analyses. Section 5.4.1 presents the specifications of the simulations, including the scenarios, the varied input parameters, and the process to create preference profiles. Afterwards, Section 5.4.2 shows the runtimes of the studied
algorithms in different settings. Considering algorithm performance, for the case
of SMT instances, i.e. preference profiles which are complete and have indiffer-

158

Resource Allocation in Social Clouds

ences, Section 5.4.3 compares the solution quality of the relevant algorithms and
evaluates how sensitive the results are with respect to certain input parameters.
Section 5.4.4 considers the same issues for SMTI instances, i.e., if preferences are
both incomplete and have indifferences.

5.4.1. Simulation Specifics
To obtain robust results and study their dependency on certain input parameters,
a systematic simulation plan is used which specifies the simulation scenarios with
the respective input parameter settings. Table 5.3 shows an overview of the most
important input parameters and the ranges of values that are used.
Table 5.3.: Simulation Input Parameters
Parameter

Range

Description

nX
nY
l
ψ
ξ

{10, 20, 50, 100, 200, 500}
{10, 20, 50, 100, 200, 500}
{0.5, 0.3, 0.1, 0.05}
{2, 5, 10}
{25, 100}

Number of requesting users
Number of providing users
Relative length of preference lists
Maximum length of ties
Percentage of users j that are correlated in the preferences of users i

The numbers of users per side, n X and nY , determine the size of the problem instance. Usually, symmetric problems with n X = nY are considered in the literature.
The evaluation, however, also considers unequally sized problem instances. The
remaining parameters are concerned with the structure of the preferences. l determines the average expected number of ranked users in the preference lists and is
used during preference creation (see parameter p D in Algorithm 5). ψ specifies the
maximum length of the ties in the preference lists. Furthermore, ξ defines if and to
what degree the preferences are correlated. For a given value of ξ, the users of the
respective other side are split into two groups, and ξ determines what percentage
of users are in the first group. This grouping of users is the same for all users of the
side, yet the subsequent randomization leads to randomized rankings within the
two groups.
For the creation of preferences, an approach similar to (Gent and Prosser, 2002;
Gelain et al., 2013) is used. The general procedure is shown in Algorithm 5. The
probability of deletion, p D , is varied between 0.5 and 0.95 to account for a decrease
in ranked users (this corresponds to the input values for parameter l). Together

5.4. PERFORMANCE OF MATCHING ALGORITHMS

159

with the 3 different values for the maximum length of ties, this parametrization
represents a full factorial 3x4 design for the variables tie-length and probability
of deletion, and allows for a study of the respective effects on the matching outcome. For all considered scenarios, 100 randomly created independent repetitions
are made.
Algorithm 5: Pseudocode of Preference Generation
Data: Deletion Probability p D = (1 − l )

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

Data: Max Length of Ties ψ
Data: Size of X and Y
Data: Correlation Parameter ξ
Result: Preference Profiles
begin
for user in set of requester or provider do
if ξ!=100 then
split users of other side into two groups of relative size ξ and 1 − ξ;
end
create random preference list;
end
for preference list of users i ∈ X do
iterate over the users in the preference list;
for user j ∈ Y do
delete j from i’s list with probability p D ;
if j is deleted, also delete i from j’s list;
end
end
for preference list do
for user in preference list do
randomly determine size of next tie length ∈ {1, ..., ψ};
set ranking of all users in the tie to the same value;
end
end
end

In case of complete preferences with indifferences (Section 5.4.3), the algorithms
DA, WO, FE, GA, TA, and GATA are considered. Finding a stable solution is trivial by using DA, hence the goal is to find a stable solution with good welfare or
fairness scores. For GA, TA, and GATA, several different objective functions are
compared. In case of welfare optimization, the main objective was to decrease the
previously introduced welfare metric. Additionally, the objective function adds a
penalty for each blocking pair, thereby discouraging unstable solutions. Similarly,
for fairness optimization the fairness metric is the main objective, and the penalty

160

Resource Allocation in Social Clouds

on blocking pairs is used as well. To show a multiple objective function, the suffix
 − EW  indicates an objective function that puts equal weight on the welfare and
fairness score, while also having a penalty on blocking pairs. Furthermore, as the
performance of DA, WO and FE depends on the way that ties are broken, these algorithms are run 50 times per preference setting to study the variability in solution
quality with respect to tie breaking. For evaluation purposes, the average, best,
and worst performance of these 50 repetitions is presented to show the variability
of results due to the tie-breaking.
For incomplete preferences with indifferences (Section 5.4.4), the considered algorithms are DA, RSMA, Király, Shift, McDermid and GSModified, as well as GA, TA
and GATA.
For the parametrization of the GA, several different parameter combination for
mutation rate, crossover rate, crossover type and population size were compared.
Similar to suggestions in the literature (see De Jong and Spears (1991)), and due
to the best performance of this combination during initial pre-tests, a population size of 50 was chosen, along with a mutation rate of 0.2, and a crossover
probability of 0.6 using a cycle crossover operator. Additionally, the best 5 solutions of a given population were always guaranteed to be ported into the next
generation. The GA was evolved for 100 evolution rounds. This is a smaller
value than usually found in the literature, yet with the trade-off between computation time and solution quality, initial experiments had shown that this is a
suitable value. As mentioned before, increasing the number of evolution rounds,
if the expected runtime permits, can help to further improve the quality of GA
and GATA. In other words, the presented results can be considered a conservative view on GA and GATA performance. For the TA, the threshold values
of {0.04 ∗ min {n X , nY } , 0.02 ∗ min {n X , nY } , 0.01 ∗ min {n X , nY } , 0} are used with
10,000 repetitions per threshold.

5.4.2. Algorithm Runtime
Table 5.1 showed that introducing indifferences or incompleteness yields NP-hard
optimization problems. The computation of an optimal solution for these problems cannot be expected in polynomial time, which is the reason why approximation and heuristic algorithms are necessary. This section gives an overview of the
runtimes of the considered algorithms in different scenarios.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

161

Finding Optimal Solutions The integer program as shown in Section 5.3.1 solves
the matching problem (with indifferences) optimally for certain cases, for example
finding a welfare-best or a maximum-size stable match. However, the calculation
of the solution to the integer program might not always be feasible in reasonable
time. In order to test the feasibility of calculating the optimal solution, the integer program was implemented for the welfare-optimal (for the SMT case) as well
as maximum-size stable match (for the SMTI case) using ILOG CPLEX 12.116 and
the corresponding Java integration to make it callable from within the simulation
tool. The solution was calculated for various problem instances, using up to 16
parallel threads on 2 Quad-Core Xeon processor with 2.53 GHz and 24-48 GB main
memory.
The runtime experiments yield the following interesting results. First, for random
uncorrelated preferences the solver is able to calculate an optimal solution for the
given problems up to problem sizes of 200x200 users in a matter of minutes. For
larger problem instances the solver usually runs out of memory. Second, correlated
preferences significantly increase the runtime of the solver. For example, for ξ = 25
the runtime of the solver increases to minutes and hours even for problem instances
of 50x50. For bigger problem instances, the solver either runs out of memory again
or does not yield a solution within several hours.
In real Social Cloud scenarios, the preference profiles can be quite diverse and include incompleteness and indifferences. Hence, the computation of an optimal
solution through integer program solvers is not feasible for larger problem instances.

Calculating Approximate or Heuristic Solutions From a computational complexity perspective, the DA is the fastest algorithm with O(n2 ). Most of the approximation algorithms for the case of incomplete preferences are similarly fast, having
runtimes of O(mn2 ), where m is an algorithm-specific factor.17 The WO algorithm,
provided that indifferences are broken first, has a runtime of O(n4 ).18 Considering
the studied heuristic algorithms, the TA also has a runtime of O(mn2 ). In this case,
m is the number of changes to the solutions (number of thresholds multiplied with
16 http://www-01.ibm.com/software/commerce/optimization/cplex-optimizer/

–
last accessed May 2014
17 Shift is the only considered approximation algorithm with a different runtime. Given that the
maximum length of ties is ψ, its runtime is O(ψ2 n2 ).
18 As mentioned before, this can be improved to an O ( n2.5 log n) bound by Feder (1992).

162

Resource Allocation in Social Clouds

the repetitions per threshold). Similarly, the GA is O(mpn2 ), where m is the number of evolution rounds and p the size of the population. For both heuristics, the
factor n2 stems from calculating the stability of the current solution, for which no
algorithm is known to have worst case time complexity better than O(n2 ). Overall,
the actual runtime then seems to depend on the algorithm-specific factor m.
Table 5.4.: Comparison of Algorithm Runtime
Algorithm

10x10

20x20

DA
WO
FE
GA
TA
GATA

<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
2 sec

RSMA
Király
McDermid
GSModified
Shift
GA
TA
GATA

<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
1 sec

50x50

Users
100x100

200x200

500x500

Complete Preferences
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
2 sec
7 sec
18 sec
2 sec
9 sec
25 sec
4 sec
15 sec
43 sec

<1 sec
1 sec
1 sec
58 sec
2 min
3 min

<1 sec
34 sec
35 sec
8 min
15 min
23 min

Incomplete Preferences
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
1 sec
3 sec
6 sec
1 sec
5 sec
8 sec
3 sec
8 sec
15 sec

<1 sec
<1 sec
<1 sec
<1 sec
<1 sec
12 sec
15 sec
31 sec

5 sec
1 sec
2 sec
3 sec
1 min
1 min
2 min
5 min

Table 5.4 shows the average runtimes of the previously introduced approximation
and heuristic algorithms for different (uncorrelated) preferences.19 The table shows
that for problem sizes up to 500 users on each side, the DA, WO, and approximation
algorithms take less than or several seconds. GA, TA and GATA take longer, on
average several seconds to compute. This is due to the large number of evolution
steps which are an essential part of the heuristics. However, even for large problem
sets finding a solution only takes a few seconds to a few minutes.
Overall, it can be seen that for most problem instances, runtimes of all the studied algorithms are in the range of seconds to minutes, which (especially for larger
problem sizes) is more than acceptable considering the NP-hard optimization problem.
19 In

general, runtimes for correlated preferences seem to be slightly higher than for uncorrelated
preferences.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

163

5.4.3. Complete Preferences With Indifferences
If the preference profiles of users are complete yet contain indifferences, the same
algorithms as in case of strict preferences can be applied by breaking the ties first.
However, as shown in Table 5.1 applying the algorithms that are optimal in the case
of strict preferences does not guarantee solution optimality anymore, and finding a
welfare- or fairness-optimal stable solution is NP-hard. This scenario is the focus of
this section, which is an extended version of (Haas et al., 2013). As the considered
algorithms always yield a stable solution of maximum size in this case, it concentrates on finding a stable solution with good welfare or fairness characteristics.
Considering GA and GATA heuristics, the subscript “− DA” indicates that they are
initialized only with DA solutions, whereas “− MIXED” means that a randomly
created mixture of DA, WO and FE solutions are used. Note that both WO and
FE were developed for symmetric problem sizes, and some of their routines in
calculating the solution do not easily extend to non-symmetric settings. Hence,
this case is not considered in this evaluation. At the same time, this shows the
advantage of having flexible heuristics to calculate matching solutions: They can
easily be applied to non-symmetric settings as well.
Unless stated otherwise, the statistical tests refer to non-parametric Wilcoxon
signed-rank tests with Bonferoni p-value adjustment in order to account for multiple comparisons. The test is used due to the simulation design (algorithms have the
same preference lists as input, leading to a paired design), as well as non-normally
distributed data. The following figures refer to the average values over the 100 independent repetitions. As DA, WO and FE were run 50 times per repetition, additional bars indicate the average, best, and worst solution out of these 50 repetitions
to study the dependency of algorithm performance on tie breaking.

Optimizing Stability and Welfare
The combination of stability and welfare maximization aims to find a stable solution with good welfare properties, i.e., where the average rank of the matched
users is as close to the respective most preferred option as possible. As mentioned
before, for SMT instances this is an NP-hard problem (and even hard to approximate), which means that finding a good solution is far from trivial. Hence, heuristics such as GA and TA might be able to yield better results than applying the algo-

164

Resource Allocation in Social Clouds

rithms developed for SM instances. The section only presents the most important
results, detailed results and additional scenarios can be found in Tables C.1 and C.3
in Appendix C. For GA, TA, and GATA, the objective of welfare maximization with
penalty on blocking pairs is used, which yields completely stable solutions for the
given scenarios.

30.00

AverageRankingofUsers

25.00

20.00

DA
TA

15.00

GATAͲDA
WO

10.00
GAͲMIXED
GATAͲMIXED

5.00

0.00
10x10

20x20

50x50
NumberUsers

100x100

200x200

Figure 5.1.: Comparison of Welfare Performance for Complete Preferences

Uncorrelated Preferences Figure 5.1 shows the results for the different algorithms for problem sizes between 10x10 and 200x200 users, averaged over different
tie-lengths. The results show that the average welfare becomes worse with an increasing number of participating users (indicated by an increasing score). This is
not surprising, as an increase in users also increases the list of potential competitors
for high-ranked users on the preference lists, thereby lowering the chance of being
matched to a high ranked user.
Considering the performance of the algorithms, GA and GATA with mixed initial
solutions perform best. Both are able to significantly increase the DA and average
WO solutions, and the solutions are slightly better than the best WO solutions (at
the level p < 0.001). This indicates that both GA and GATA are able to improve
upon WO solutions and yield a superior solution quality. Additionally, the fairness
scores of the GA and GATA solutions are on average slightly better than the WO
(average and best) solutions.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

165

Considering the GA performance with different initial solutions, the results show
that the GA with only DA solutions is able to significantly improve upon the DA
solutions, and also yields better solutions than WO for problem sizes up to 50x50.
For larger number of users, the WO might yield better solutions, yet the increase
in solution quality compared to the initial solutions is still significant (p < 0.001).
If GA and GATA use mixed initial solutions, the heuristics are able to improve
upon them as well. This shows two issues: First, the GA effectively improves the
given starting solutions, showing its usefulness. Second, the quality of the starting
solutions determine the quality of the heuristic solutions, which means that feeding
the heuristics with promising initial solutions increases their performance.
Interestingly, the TA does not perform as well. In the given setting, the TA starts
with a DA solution and then tries to find a better solution as specified in Section
5.3.3. Figure 5.1 shows that only for small problem instances the TA is able to significantly improve the solution quality of the starting DA solution. For larger problem instances, TA performs considerably worse compared to WO and GA. This
indicates that starting with DA solutions is not promising for TAs, which might get
stuck in local optima. In contrast, the GA with its ability to sample large search
spaces is an adequate heuristic for SMT instances. Due to this, the performance of
GA and GATA are basically the same as TA is not able to significantly improve the
given solution.

Effect of Preference Correlation In this scenario, preferences are correlated with
the factor ξ = 25, meaning that the set of user IDs is split into two sets of relative
size 25% and 75%, and the 25% highest ranked users in all the preference profiles
(of each side) are drawn from the same set.
Considering finding a welfare-optimal stable solution, Figure 5.2 compares the performance of DA, WO, GA, TA, and GATA. The results are qualitatively very similar
to the case of uncorrelated preferences. In general, the results show that the average welfare decreases compared to the uncorrelated case, which is not surprising
as the correlation means that some users are more likely to be matched with lowerranking users as there is an increased competition for higher-ranked users. The relative ranking of the algorithm performance is practically the same. GA and GATA
perform significantly better than the other algorithms on average (p < 0.001), and
only the best WO solution yields comparable results (for all considered scenarios,
the signed-rank test reveals a statistical difference between the best WO solution

166

Resource Allocation in Social Clouds
30.00

AverageRankingofUsers

25.00

20.00

DA
TA

15.00

GATAͲDA
WO

10.00
GAͲMIXED
GATAͲMIXED

5.00

0.00
10x10

20x20
50x50
NumberUsers

100x100

Figure 5.2.: Comparison of Welfare Performance for Complete and Correlated Preferences

and GA (GATA), although the absolute difference is negligible). Another interesting finding is that the spread in solution quality is smaller for correlated preferences, i.e., is less dependent on the way that ties are broken. This is indicated by
the smaller range of best and worst solutions averaged over the 100 repetitions.
Sensitivity analyses were also performed on varying values for maximum tie
lengths. The results indicate that the relative performance of GA and GATA is
better than WO particularly for small tie lengths. However, the qualitative results
are similar.

Optimizing Stability and Fairness
Besides welfare optimization, finding a stable and particularly fair solution is another common goal in two-sided matching. However, even for strict preferences
this is an NP-hard problem, and the approximation algorithm FE can be applied in
SM and SMT instances by appropriately breaking ties. This part of the evaluation
aims to study the performance of the DA, FE, GA, TA, and GATA algorithms for
finding stable and fair solutions. As before, it concentrates on the most important
findings, and the detailed results can be found in Tables C.2 and C.4 in Appendix
C. For GA, TA, and GATA, the objective of fairness maximization with penalty on
blocking pairs is used, which yields completely stable solutions for the given scenarios.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

167

30.00

AverageDifferenceinRanking

25.00

20.00

DA
TA

15.00

GATAͲDA
FE

10.00
GAͲMIXED
GATAͲMIXED

5.00

0.00
10x10

20x20

50x50
100x100
NumberUsers

200x200

Figure 5.3.: Comparison of Fairness Performance for Complete Preferences

Uncorrelated Preferences Figure 5.3 presents the fairness performance of the
considered algorithms. As expected from theory, the DA has the worst performance of the considered algorithms. Furthermore, as before TA finds improvements (in this case: fair solutions) only for small problem instances. For larger
instances the TA performance considerably decreases. The figure also shows that
the range of best and worst solutions can be considerably high, which indicates
that especially for DA tie-breaking greatly affects the solution quality.
Comparing the performance of DA, FE and GA the results indicate that the GA
and GATA yield significantly better results with respect to fairness if it is initialized with mixed solutions, and also outperforms the average FE solution. This is
confirmed by Wilcoxon signed-rank tests at the level p < 0.001. Over all considered
scenarios, the signed-rank test indicates a significant difference between the best FE
solution and the GA solution with mixed initial solutions, although the practical
effect size can be considered negligible. The performance of GATA-DA, initialized
only with DA solutions, is significantly better than DA solutions (p < 0.001), yet,
for all but very small problem sizes up to 20x20 users, it is worse than the average
FE solution. Considering sensitivity with respect to input parameters, TA performs
considerably better for larger tie lengths (see Tables C.2 and C.4 in Appendix C).
Another interesting result is the performance of the GA with equal weight on welfare and fairness. Tables C.1 - C.4 show that with this objective function, the fair-

168

Resource Allocation in Social Clouds

ness of the solutions can be substantially increased while resulting only in a small
decrease in welfare performance. This points out one of the main advantages of using heuristics to solve two-sided matching problems: the ability to optimize several
metrics simultaneously.

8.00

AverageDifferenceinRanking

7.00
6.00
DA

5.00

TA
4.00

GATAͲDA

3.00

FE
GAͲMIXED

2.00

GATAͲMIXED
1.00
0.00
10x10

20x20
50x50
NumberUsers

100x100

Figure 5.4.: Comparison of Fairness Performance for Complete and Correlated Preferences

Effect of Preference Correlation The results for correlated preferences are similar. Figure 5.4 shows that GA and GATA with fairness optimization and penalty
for unstable pairs is able to yield almost perfectly fair solutions for the studied
scenarios. The performance of GA and GATA with mixed initial solutions are significantly better with respect to fairness, as a Wilcoxon signed-rank test at the level
p = 0.01 reveals. Furthermore, the TA performs slightly better for correlated preferences, and is able to provide comparably good results for problem sizes up to
50x50 users.

5.4.4. Incomplete Preferences With Indifferences
In the most general case, preference profiles can be both incomplete and contain
indifferences, in which case the common goal of matching algorithms is to find a

5.4. PERFORMANCE OF MATCHING ALGORITHMS

169

stable solution of maximum size. As shown in Table 5.1, there are several approximation algorithms and heuristics specifically developed for this case. The focus
of this section is to study how GA, TA, and GATA heuristics perform against these
specialized algorithms in different scenarios. For statistical evaluation, Wilcoxon
signed-rank tests are used for the same reasons as in the previous section.

Comparing Solution Quality
To compare the overall efficiency of the algorithms, this section studies the ability
of the mechanisms to compute a stable allocation of maximum size. The following
figures represent the average values based on the 12*100 independent repetitions
for each scenario, a more detailed analysis of the effects of preference structures on
the results is presented subsequently. Similar to the previous section, for GATA the
suffix  − MIXED  indicates that initial solutions are based on DA solutions plus
an additional solution from applying Király, McDermid, and GSModified, respectively.20 If no suffix is provided, GA and GATA are initialized only with DA solutions. Considering the optimization function, GA, TA, and GATA use a weighted
function that emphasizes the number of matched pairs as main goal, and also tries
to optimize welfare as secondary goal. More specifically, the goal is to minimize
an objective function of the form ( Max (n X , nY ))2 ∗ ( Min (n X , nY ) − NumPairs) +
Wel f are. This ensures that the number of matched pairs is the dominant objective,
yet still allows for the simultaneous optimization of welfare. For the evaluation and
comparison with the approximation algorithm, such an optimization function represents a pessimistic comparison as the welfare scores tend to increase with a larger
number of matched pairs, indicating that welfare and matched pairs are conflicting
optimization goals.
Performance in this case can be measured in two ways. On one hand, the percentage of matched users can be considered as a proxy for algorithm performance.
Depending on the preferences, however, it cannot be guaranteed that there is a stable solution where all users are matched. This suggests measuring the performance
relative to the optimal solution as an indicator how close the algorithms are to the
optimal outcome.

20 As seen in Section 5.4.2, the runtime for the approximation algorithms is comparably short, hence

using them in addition to the DA to get initial solutions does not increase the relative runtime
of GA or GATA.

170

Resource Allocation in Social Clouds
1

PercentageofMatchedUsers

0.95
0.9

DA
GSModified

0.85

McDermid
RSMA

0.8

Shift
TA

0.75

GATA
GA

0.7

Király
GATAͲMIXED

0.65
0.6
10x10

20x20

50x50

100x100
ProblemSize

200x200

500x500

Figure 5.5.: Algorithm Performance in Number of Matched Pairs, Uncorrelated Preferences

Figure 5.5 displays the results for the first performance measure, the percentage of
matched users. Averaged over the different parameter values for tie length and
length of preference lists, the results show that with an increasing number of users,
the percentage of matched users in the solutions increases as well. Especially for
large problem sizes, the majority of users are matched (the next section shows that
this seems to depend on the lengths of the preference lists). There are apparently
considerable differences in the performance of the algorithms. DA performs worst,
which is not surprising as it was not specifically developed for this scenario. Applying the RSMA improvement cycle on the DA solution is able to significantly
increase the number of matched pairs. Considering the approximation algorithms
described in Section 5.3.2, the Király algorithm performs best, while GSModified,
McDermid and Shift are slightly worse. Considering the Shift algorithm, the extended results in Table C.5 in Appendix C show that it yields its best performance
in case the maximum length of ties is 2, which is the core scenario for which it was
developed. However, the Király algorithm usually provides better solutions, and
for larger tie lengths the performance of Shift decreases considerably.
Both the GA and TA perform very well, being slightly outperformed only by Király.
This indicates that in contrast to the previous scenario with complete preferences,
TA is a useful heuristic in SMTI scenarios. Additionally, the results for GATA and
GATA-MIXED show that instantiating the GATA with solutions from approximation algorithms can significantly increase the solution quality of GATA. In par-

5.4. PERFORMANCE OF MATCHING ALGORITHMS

171

RelativePerformancetoOptimalSolution

ticular, GATA-MIXED on average outperforms even the best approximation algorithm, Király, in the considered scenarios. The statistical analysis with a Wilcoxon
signed-rank test reveals that there is no significant difference between GA, GATA
and Király, and that GA, GATA and Király are better than TA, Shift, McDermid,
GSModified as well as RSMA (p < 0.001). Furthermore, GATA-MIXED yields better solutions than GA, GATA, and Király (p < 0.001).
1
0.99
0.98
0.97
0.96
0.95
0.94

Figure 5.6.: Algorithm Performance Relative to Optimum, Uncorrelated Preferences, 10x10
to 100x100 users

Next, Figure 5.6 shows the relative performance of the considered algorithms with
respect to the optimum solution calculated with the Integer Program formulation
as shown in Section 5.3.1. Due to runtime and memory considerations, only the
problem sizes with 10 to 100 users on each side are considered here. The graph
shows that for all considered scenarios the algorithms are within 95% of the optimum solution, i.e., are able to find relatively large matches which are close to the
optimal solution.
Table 5.5.: Algorithm Performance Relative to Optimal Solution, Incomplete Preferences,
10-100 Users per Side
Statistic

DA

GSMod.

McDer.

RSMA Shift

TA

GATA GA

Király GATAMIXED

Mean
Median
S.D.
Min
Max
Opt.
calc.
(%)

0.952
0.965
0.051
0.500
1.000
0.228

0.980
1.000
0.032
0.778
1.000
0.619

0.983
1.000
0.030
0.667
1.000
0.660

0.983
1.000
0.031
0.667
1.000
0.669

0.989
1.000
0.023
0.750
1.000
0.731

0.991
1.000
0.022
0.667
1.000
0.763

0.991
1.000
0.019
0.857
1.000
0.754

0.984
1.000
0.029
0.750
1.000
0.684

0.991
1.000
0.020
0.750
1.000
0.759

0.995
1.000
0.014
0.857
1.000
0.813

172

Resource Allocation in Social Clouds

Table 5.5 presents several statistics for the relative performance of the algorithms
compared to the optimal solution. Interestingly, the heuristics not only perform
well on average, they also calculate the optimal solution in more cases than the
approximation algorithms. Furthermore, the standard deviation is lower as well.
As approximation algorithms provide a guaranteed quality bound, i.e., worst case
performance, Table 5.5 shows that the worst-case performance is best for Király
and GATA-MIXED.21
Overall, similar to the previous findings about the percentage of matched users,
the consideration of average and worst-case performance relative to the optimal
solution shows that GATA-MIXED not only yields the best solutions, but also the
solutions with the best worst-case performance. This is particularly interesting as
it indicates that for the considered scenarios, the practical quality bounds of GA,
GATA, and especially GATA-MIXED are comparable (or even better) to the quality
bounds of the approximation algorithms. However, as heuristics are not able to
provide definite quality bounds, there might be scenarios where the worst-case
performance is lower than that of the approximation algorithms. Nevertheless, the
performance of the studied heuristics is more than promising.
Table 5.6.: Algorithm Performance for Welfare and Fairness Relative to GATA-MIXED, Incomplete Preferences. Percentages indicate to what degree the GATA-MIXED solution improves upon the respective algorithm.
Size

DA

RSMA

10x10
20x20
50x50
100x100
200x200
500x500

10.2%
13.9%
13.7%
12.2%
13.4%
18.5%

5.4%
13.6%
31.1%
52.1%
79.9%
125.2%

10x10
20x20
50x50
100x100
200x200
500x500

2.3%
3.8%
14.1%
35.5%
72.9%
75.5%

4.7%
21.4%
89.5%
196.9%
338.0%
378.2%

21 The

GSMod.

McDer.

Shift

Király

TA

GA

GATA

Welfare Score Increase relative to GATA-MIXED
5.7%
6.3%
1.1%
3.9%
0.0%
11.3%
14.5%
2.5%
7.1%
0.5%
14.9%
26.6%
3.4%
8.7%
4.9%
16.7%
38.5%
3.0%
8.7%
7.6%
20.0%
50.0%
12.5%
11.6%
11.8%
26.2%
60.9%
18.5%
18.3%
17.9%

0.9%
1.1%
1.1%
1.0%
1.0%
1.2%

0.4%
0.8%
0.8%
0.9%
1.1%
1.3%

Fairness Score Increase relative to GATA-MIXED
3.1%
5.1%
0.4%
2.5%
-0.2%
9.3%
19.8%
0.2%
4.7%
-0.7%
30.3%
68.8%
2.1%
14.8%
7.3%
63.6%
145.2% 4.7%
36.3%
30.5%
108.3% 224.7% 70.6%
72.4%
69.9%
102.6% 203.4% 75.9%
76.6%
73.3%

0.3%
0.0%
0.7%
0.7%
4.5%
4.9%

0.1%
0.0%
0.8%
1.8%
5.1%
5.0%

worst-case performance of 0.66 for McDermid is exactly the guaranteed 3/2 approximation
factor of the considered approximation algorithm.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

173

Besides the superior performance in the number of matched pairs, it is also interesting to consider the performance of the algorithms with respect to welfare and
fairness characteristics, even though the approximation algorithms are not specifically developed for this combination of metrics. Table 5.6 presents the average
welfare and fairness performance relative to GATA-MIXED. For example, a value
of 50% indicates that the welfare score of an algorithm was 50% higher (i.e., worse)
than the welfare of the GATA-MIXED solution. The results show that not only does
GATA-MIXED yield better solutions with respect to the number of matched pairs,
the solution quality with respect to welfare and fairness is considerably better than
the approximation algorithms.22 The relative improvements in welfare and fairness are particularly high for larger problem instances, which indicates that the respective solution quality can be considerably increased by the use of the proposed
heuristics. For example, compared to the best approximation algorithm Király, welfare improvements of up to 18% and fairness improvements of up to 76% can be
achieved, which means that users are on average matched to higher ranked partners, and that the solution treats both sides more equally. Table 5.6 also shows that
RSMA, which focuses on finding improvements for users of one market side, yields
comparably unfair solutions, and potential fairness and welfare improvements by
using the heuristics can be substantial. These results show that the heuristics not
only provide good solutions for the standard combinations of goals, but also for
multiple objectives such as finding a maximum size stable match with good welfare or fairness properties. This can be especially important in social settings where
such welfare and fairness aspects are of importance, making the proposed heuristics particularly beneficial in these settings.

Effect of Preference Structures and User Distributions
To study the sensitivity of the results on different design parameters, this section
looks at the algorithm performance in case of different tie lengths, different preference lengths, and correlation in preferences.

Influence of Tie Lengths For the different maximum lengths of ties, Figure 5.7
shows that, as before, GA and TA perform quite well. In most cases their perfor22 Interestingly,

TA seems to find slightly better solutions for small problem instances. The relative
improvement, however, is small, and taking into account the solution quality for larger problem
instances, GATA-MIXED provides the overall best solution quality.

174

Resource Allocation in Social Clouds
1

0.99

RelativePerformance

DA
RSMA

0.98

GSModified
McDermid
0.97

Shift
Király

0.96

TA
GA

0.95

GATA
GATAͲMIXED

0.94
2

5
MaximumTieLength

10

Figure 5.7.: Performance of Matching Algorithms, Random Preferences

mance is better than the approximation algorithms, and only Király yields better
solutions. GATA-MIXED yields consistently the best performance in the considered scenarios. Overall, the relative performance differences become smaller with
increasing tie lengths. These observations are confirmed by a statistical analysis
applying Wilcoxon signed-rank tests. For small maximum tie lengths, ψ = 2 and
ψ = 5, GA and Király are significantly better than TA, and GATA-MIXED is significantly better than GA and Király (p < 0.001). For ψ = 10, the performance of the
algorithms is more similar overall. TA is in fact slightly better than GA (p = 0.001)
and yields similar solutions than Király (no statistically significant difference), and
GATA-MIXED is still better than Király at the level p = 0.001.
Influence of Preference Lengths Figure 5.8 shows the results for different average preference lengths, both for the percentage of matched users and relative
to the optimal solution. For larger values of l, more users will be deleted from
each others’ preference profiles, thereby shortening the profiles and increasing the
probability that some users might be unmatched in a solution. This is especially
reflected in Figure 5.8a, which shows that the percentage of matched users considerably decreases if users have short preference lists. Per definition of parameter
l, the preference length is measured proportionally to the number of users, which
means that for larger problem sizes, the absolute preference lengths increase for the
same l. This might explain the findings shown earlier in Figure 5.5, which showed
that with increasing user size the percentage of matched users increases.

5.4. PERFORMANCE OF MATCHING ALGORITHMS

175

1

PercentageofMatchedUsers

0.95
DA
0.9

RSMA
GSModified
McDermid

0.85

Shift
Király
TA

0.8

GA
GATA
0.75

GATAͲMIXED

0.7
0.5

0.3
0.1
RelativeLengthofPreferenceLists

0.05

(a) Percentage Matched Users
1

0.99

RelativePerformance

DA
RSMA

0.98

GSModified
McDermid
0.97

Shift
Király

0.96

TA
GA

0.95

GATA
GATAͲMIXED

0.94
0.5

0.3
0.1
RelativeLengthofPreferenceLists

0.05

(b) Performance Relative to Optimum
Figure 5.8.: Comparison of Algorithms for Different Preference Lengths

In more detail, Table C.5 in Appendix C shows that the absolute number of matched
pairs decreases with decreasing l, which was expected. Additionally, Figure 5.8b
shows that with increasing l the relative performance of the algorithms, compared
to the optimal solution decreases (slightly). Considering the relative performance
of the algorithms, Figure 5.8 shows that Király is the best approximation algorithm.
GA, TA, and GATA yield solutions similar to Király, yet often better than other
approximation algorithms. As before, GATA-MIXED outperforms all other algorithms. This increase in solution quality is statistically significant for smaller values
of l (at the level p = 0.01 for l = 0.1 and l = 0.05), yet relatively small. Especially for
larger values of l, i.e., longer preference lists, it seems to be easier to find large sta-

176

Resource Allocation in Social Clouds

ble matches which makes the differences between algorithms marginal. However,
when users provide the preference ranks of other users manually in larger Social
Clouds, smaller values for l are realistic as users might have time and cognitive
limitations in providing the preference ranks.

1

RelativePerformancetoOptimum

0.998
DA
GSModified

0.996

McDermid
RSMA
0.994

Shift
Király

0.992

TA
GA
GATA

0.99

GATAͲMIXED
0.988
10x50

10x100

20x50
20x100
ProblemSize

50x10

50x100

Figure 5.9.: Algorithm Performance, Asymmetric Problem Instances

Scenarios with Asymmetric Numbers of Users Figure 5.9 shows the performance of the algorithms relative to the optimal solution for problems where the
numbers of requesting and providing users are unequal. The overall efficiency of
the considered algorithms is better than in the case of symmetric problem instances.
Furthermore, the higher the difference in size, the closer the relative performance of
the algorithms. This can be expected, as the maximum number of agents matched
is bounded by the number of users of the smaller side, and the probability to match
all such users increases if (relatively) more users of the second side are present in
the market. As a result, all algorithms calculate an optimal solution in 99% of the
cases. Figure 5.9 also shows that, while the performance increase is comparably
small, the performance of the GA, GATA, and GATA-MIXED heuristics is slightly
better than the approximation algorithms, which confirms the previous findings.
In particular, GATA-MIXED is able to find an optimum in all the studied problem
instances.

RelativePerformancetoOptimalSolution

5.4. PERFORMANCE OF MATCHING ALGORITHMS

177

1
0.99
0.98
0.97
0.96
0.95
0.94

Figure 5.10.: Algorithm Performance, Correlated Preferences
Table 5.7.: Algorithm Performance Relative to Optimal Solution, Incomplete Correlated
Preferences
Statistic

DA

GSMod.

McDer.

RSMA Shift

TA

GATA GA

Király GATAMIXED

Mean
Median
S.D.
Min
Max
Opt.
calc.
(%)

0.952
0.962
0.048
0.500
1.000
0.205

0.978
1.000
0.033
0.750
1.000
0.588

0.980
1.000
0.033
0.667
1.000
0.618

0.979
1.000
0.035
0.500
1.000
0.618

0.987
1.000
0.024
0.818
1.000
0.690

0.989
1.000
0.023
0.818
1.000
0.730

0.989
1.000
0.022
0.800
1.000
0.721

0.983
1.000
0.028
0.750
1.000
0.683

0.990
1.000
0.020
0.857
1.000
0.745

0.994
1.000
0.014
0.857
1.000
0.799

Correlated Preferences For correlated preferences, Figure 5.10 shows the relative performance ranking of the algorithms over all combinations of problem size,
lengths of preference lists, and maximum lengths of ties. Similar to the scenario
with uncorrelated preferences, the heuristics perform exceptionally well. As before, GATA-MIXED yields on average the best solutions of the considered algorithms. Table 5.7 also shows that it has a low variance in solution quality, and that
its worst case solution quality is better than Király and the other approximation
algorithms. Whereas the general worst-case performance seems to be similar to
the case of uncorrelated preferences, the number of times where the algorithms are
able to find an optimal solution is lower. Compared to the findings in Table 5.5, the
percentage of scenarios where an optimal solution was found decreases from up to
86% to 80% in the best case.

178

Resource Allocation in Social Clouds

5.5. Summary
This chapter introduced the concept of preference-based resource matching and
several algorithms that are used to calculate allocations with certain desired properties. Section 5.1 provided the necessary notation and fundamental theorems, and
classified the matching problems according to preference structures and optimization metrics. Section 5.2 continued with an overview of related work, and Section
5.3 described the algorithms considered to calculate a solution for the matching
problem. Section 5.4 evaluated the performance of the proposed heuristics and
existing algorithms for various preference structures.
As the complexity to calculate such solutions is, for general preference types, NPhard, either approximation algorithms or heuristics are needed as calculating the
optimal solution might be infeasible. Heuristics provide the flexibility to consider
various combinations of objective functions. To determine their performance relative to existing algorithms, Section 5.4 compared and evaluated the performance of
the considered algorithms for several scenarios in order to address research question 2.1. For the scenario of complete preferences with indifferences, the evaluation
showed that depending on the initial solutions, a genetic algorithm (potentially
combined with a Threshold Accepting algorithm) improves upon the initial solutions, and on average yields similar or better solutions than existing algorithms. In
case of incomplete preferences, the heuristics (especially GATA with mixed initial
solutions) are able to yield better solutions on average than even the best approximation algorithms concerning the number of matched users, and this average improvement is also statistically significant. Furthermore, the solutions found by the
heuristics outperform the solutions of approximation algorithms with respect to
additional criteria such as welfare or fairness to a substantial degree. This is a particularly important result as platforms with social contexts, such as Social Clouds,
might value welfare and fairness aspects in addition to the standard set of optimization goals for which approximation algorithms exist (stability and number of
matched pairs). Users of a Social Cloud benefit from the application of the proposed heuristics as they are, on average, matched to a higher ranked partner, and
the solution is fairer to both requesting and providing users.
Overall, this section showed that heuristic algorithms provide a useful, flexible and
powerful tool to compute solutions for preference-based matching with a short
runtime. Potentially, the solution quality of the heuristics can be even further

5.5. SUMMARY

179

improved by applying subsequent optimization steps such as RSMA (requesteroptimal-stable-matching). Initial results indicate that this can be the case for some
repetitions and scenarios. A conclusive study of the combination of the heuristics
with existing algorithms (such as improvement cycles), thus, is an interesting area
of further research. Another interesting aspect is the computational optimization
of the algorithms. Especially GAs are considered suitable for parallelization, which
might decrease their runtime and make them even more useful.
In the context of preference-based resource exchange in social contexts, the results
are especially promising. The actual structure of user preferences can be diverse
and likely includes incompleteness and indifferences, especially when users have
larger social networks and cannot provide a definite rank for each single user (or do
not want to share with certain users). For example, in the Social Cloud prototype
(Section 2.1.4), the preference interface allows for the specification of “blocked”
users with whom sharing is not wanted, and providing the same rank for several
users indicates indifference. The absence of an explicitly provided rank can either be interpreted as being indifferent or also not wanting to be matched to these
users. The latter decision will most likely affect the allocation, as not considering
many users in the preference ranks leads to short preference lists. The heuristics
are, however, well suited to handle such situations, and provide solutions that are
on average closer to the optimal solution than the considered approximation algorithms (see Section 5.4.4).
Having identified that heuristics provide similar or superior performance to existing algorithms and are adaptable to various different scenarios, the next chapter
considers two case studies on preference-based matching. The first case study considers the effects of strategic preference manipulation for the different algorithms,
both from a system and a user perspective. Especially for participating users, this
can be interesting as it possibly provides strategic guidelines of how to act in such
a resource allocation market. The second case study considers the case of dynamic
allocation of resources. In such a scenario, resource supply and demand is not
matched in distinct, batch-like allocations, but continuously over time. As a frequent recalculation of the allocation, or migration of allocated resources might not
be possible, the case study considers heuristics to cope with such dynamic supply
and demand.

Chapter 6.
Incentive Compatibility and Dynamic
Allocation
“It is difficult to advise participants in markets that use stable matching mechanisms
when to behave straightforwardly (i.e. in a way that reveals their true preferences) and
when there might be opportunities to behave strategically, and if so, how.”
(Roth and Rothblum, 1999)

T

HE solution quality of two-sided matching problems, which was the focus of
the last chapter, is an important consideration in preference-based resource
allocation. Yet, there are other aspects that might be of interest when a two-sided
matching approach is applied in practice. This chapter looks at two of these aspects,
namely the manipulation of preferences and the dynamic allocation of resources.
The first aspect considered in this chapter is preference manipulation. The matching
mechanisms studied in the last chapter all require the preference profiles of the
participating users as input for their calculations, and the solution of the respective
mechanisms depends on the profiles submitted by the users. An important observation in this case is that, in general, the submitted preference profiles might not be
the true preference profiles of the users. Under the assumption that users want to
be matched with a partner as highly ranked as possible, rational users might have
an incentive to manipulate the submitted preference profiles in order to increase
their chance of being matched with a higher-ranked partner.
This is the focus of Section 6.1, which considers the existence and the effects of
preference manipulations for the mechanisms introduced in the previous chapter.

181

182

Incentive Compatibility and Dynamic Allocation

In particular, two aspects are considered: 1) the degree to which users can benefit
(or be worse off) by manipulating their preferences; and 2) the effect of preference
manipulation on the solution quality of different matching mechanisms.
As the calculation of a solution depends on the submitted preference profiles of all
participating users, preference manipulation does not only affect the manipulating
users, but other users as well. Furthermore, the matches calculated by the algorithms might be stable for the submitted preference profiles, but not for the true
profiles of the users. The introduction of instability, considering the true profiles
of the manipulating users, is therefore also considered as potential manipulation
effect.
Besides studying the effects of preference manipulation on the market, for individual participants it might be interesting to know under which circumstances preference manipulation is useful, and what manipulation strategies are most promising. Due to the complexity of the matching problem and the multitude of potential
manipulation types, straightforward strategy recommendations for participants in
different markets do not exist, and existing results are limited to special circumstances (Roth and Rothblum, 1999). Another interesting issue is how robust the
different mechanisms are against strategic manipulation. These two aspects are the
focus of Section 6.1.3.
In the second part of this chapter, Section 6.2 considers the application of two-sided
matching mechanisms in dynamic scenarios, using the prototype of a Social Compute Cloud (see Section 2.1.4). Matching mechanisms are usually applied in static
settings, where users submit their preference profiles once or only at certain time
intervals, and the mechanism calculates a solution for the given submitted profiles.
In (social) resource exchange settings such as a Social Cloud, the allocation can be
more dynamic in the sense that users might enter or leave the market in between
the time intervals. This creates new, unallocated supply and/or demand (either by
arriving users, or by freeing allocated resource offers/requests), which is denoted
as intermediate supply and/or demand for the rest of this chapter. In such a case, the
question arises how such intermediate supply and demand can be handled.
For technical and computational reasons, it might not be feasible to recompute the
entire allocation every time a user enters or drops out. For example, users in existing matched pairs might have agreed that the allocation is valid for a certain
time, and a reallocation might break up such a match. Hence, the second part

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

183

of this chapter looks at the effects of dynamic supply and demand in preferencebased resource allocation. In particular, Section 6.2 considers potential algorithms
to match supply and demand, and investigates the effects of different algorithms
for dynamic allocation on the market.
The chapter is structured as follows. Section 6.1 introduces the concept of preference manipulation and studies its effects on allocations in preference-based resource matching. Section 6.2 discusses and evaluates different algorithms to match
intermediary supply and demand. Section 6.3 summarizes the findings and gives
an outlook on potential next steps.

6.1. Strategic Manipulation in Resource Allocation
In centralized resource allocation mechanisms, the allocation is determined based
on the information submitted by the participants. In general, this submitted information does not need to be truthful, i.e., it might not reflect the true valuation
or preference of a participant. Depending on the mechanism, a participant might
benefit from sending untruthful information, which in turn could lead to an allocation that does not reflect the true preferences in the market. Untruthful information about preferences can have serious implications in the real world, for
example schools ranked low in the school choice problem have been closed due
to the ranking (see Abdulkadiroğlu et al. (2009)). Such behavior can also lead to
the emergence of blocking pairs with respect to the true preferences, i.e., if an allocation is calculated based on submitted preferences it does not necessarily have
the same properties, such as stability, as under the true preferences. This section
considers this issue in the context of preference-based two-sided matching. Specifically, after introducing the necessary concepts and theoretical results, it studies the
potential gains and losses of manipulating users, and compares the robustness of
the considered matching mechanism against such strategic manipulation.

6.1.1. Theoretical Results and Manipulation Strategies
In the economic field of mechanism design, impossibility theorems provide guidance about which combinations of market goals can be achieved (see e.g. Parkes
(2001) for an overview), and incentive compatibility is one of the most frequently

184

Incentive Compatibility and Dynamic Allocation

considered properties of a mechanism. Participants have private information about
their true preferences, and reveal parts of this information through interaction with
the mechanism (e.g., through bids or submission of preference rankings). However, depending on the market mechanism it might not be best for them to reveal
their true private information. The aspect of incentive compatibility in this case is
defined as follows:
Definition 8 ( Incentive Compatibility in Two-Sided Matching ). A two-sided
matching mechanism is incentive compatible if submitting the true preference profile is
the best strategy in equilibrium. If this applies for dominant strategies, then the mechanism
is said to be strategy-proof.
Note that incentive compatibility in general is defined for any equilibrium concept, such as Nash, Bayes, or dominant strategy equilibrium. In the context of
two-sided matching, previous work has focused on dominant strategies, in which
case strategy-proof would be the correct term. Adopting the standard of previous
literature in this field, unless otherwise stated incentive compatibility will refer to
dominant strategies in the subsequent considerations.
From the viewpoint of a market designer, achieving incentive compatibility (ideally
in dominant strategies) is important for several reasons. On one hand, participants
do not have to calculate complex strategies of how to act on a market, as it is in
their best interest to act based on their true preferences. On the other hand, it guarantees that the solution quality calculated with submitted preferences is reflecting
the true quality, e.g., stable solutions might not necessarily be stable if the private
preferences differ from the submitted ones.
In the case of two-sided matching where participants have private information
about their true preferences, incentive compatibility studies whether it is best for
participants to reveal these true preferences while acting on the market or not. The
fundamental result considering incentive compatibility in two-sided matching was
developed by Roth (1982) and summarized in Roth and Sotomayor (1992):
Theorem 4 ( Roth, 1982 ). No stable matching mechanism exists for which stating the
true preferences is a dominant strategy for every agent.

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

185

This result has serious implications on the design of a two-sided matching mechanism. As stability is commonly (and empirically) considered the most important
property of a two-sided matching mechanism (see e.g. Roth (2008)), incentive compatibility needs to be sacrificed if stability is to be guaranteed. Furthermore, Alcalde and Barberà (1994) show that strategy proofness is also incompatible with
individual rationality and Pareto efficiency. Considering the DA, Roth and Sotomayor (1992) also show that the DA is strategy-proof for the proposing side, yet
not the accepting side.
Although this seems to be a rather negative result, the implications of the impossibility of incentive compatibility for all participants are less clear. As Roth and
Rothblum (1999) noted: “However the existing theoretical results do not generally
allow us to address the considerable demand for practical advice about how to
participate in such markets, once they are established. It is difficult to advise participants in markets that use stable matching mechanisms when to behave straightforwardly (i.e. in a way that reveals their true preferences) and when there might be
opportunities to behave strategically, and if so, how. This also suggests that there
are some gaps in our understanding of why stable matching mechanisms work so
well in practice” (Roth and Rothblum, 1999, p.21). For example, as one of the few
results in the literature, Roth and Rothblum (1999) find that only a small number
of participants have incentives to reveal altered preference rankings. Additionally,
Pini et al. (2011a) show that there can be non-strategy-proof matching mechanisms
which are NP-hard to manipulate, i.e., finding a successful manipulation strategy
can be a hard problem.

Related Work
Due to its practical relevance, e.g. in school choice and college-admission markets,
preference manipulation has been subject of many studies. Most of these focus on
preference manipulation either in the classic DA, or in the adapted mechanism for
many-to-one matchings.
Considering its strategic properties, the DA is strategy-proof for the proposing
side, yet not strategy proof for the accepting side (Roth and Sotomayor, 1992, p.
90). Moreover, not putting its most preferred alternative first is a dominated strategy for the users of the accepting side (Roth and Sotomayor, 1992, p. 105). Abdulkadiroğlu et al. (2009) show that (again in the case of strict preferences) for any

186

Incentive Compatibility and Dynamic Allocation

tie breaking rule, there is no mechanism that is strategy-proof and dominates the
DA. Ashlagi and Klijn (2012) consider manipulation in the DA and show that all
weakly beneficial group manipulation strategies of accepting users are beneficial
for all other accepting users and harmful for all proposing users. Furthermore, this
is true if users from the accepting side apply a truncation of preferences. Studying
the prevalence of manipulation in experimental settings, Echenique et al. (2009)
show that truncation of preferences for the accepting side in a DA is applied only
rarely.
In many-to-one settings, several studies looked at the effects of manipulation in the
respective markets. Under certain conditions, the percentage of users that can successfully manipulate their preferences in a student-optimal stable matching converges to zero for large markets (Kojima and Pathak, 2009; Lee, 2011). Abdulkadiroğlu et al. (2009) discuss the effect of strategy-proofness on efficiency (measured
in average rankings) in the school-choice problem using date from New York and
Boston school districts. Furthermore, Kesten (2012) studies manipulation strategies
in the school choice problems. In particular, he considers the option for schools
to manipulate their submitted capacity (i.e., offering less capacity than available),
and the possibility to pre-arrange matches in which case the involved student does
not participate in the actual matching procedure. Whereas some of the studied
mechanisms are immune to capacity manipulation, Kesten shows that all studied
mechanisms are not immune to pre-arranged matches.

Fixed Strategies
In order to study the implications of a non-achievable incentive compatibility for
different matching algorithms, the manipulation strategies need to be defined.
Based on literature in this field, several strategy types are considered for the subsequent evaluation.

Re-ordering Matching algorithms cannot guarantee that each participant is
matched to its most preferred alternative. Hence, one potential reasoning for this
strategy is that if participants are not likely to be matched with their most preferred alternative, putting more preferred alternatives (in their true preferences) in
lower ranks might result in a better match for them. For example, if a participant
is on average matched to its third choice, putting their true first choice at rank 3

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

187

might yield a better result for them. However, as the matching depends on the
preferences of the other participants, it is not straightforward to see whether such
a strategy might be useful. Furthermore, for strict and complete preferences Roth
and Sotomayor (1992) show that not putting the most preferred alternative first is
a dominated strategy for the DA. The re-ordering or shuffle strategy applies this
reasoning to create a new preference ranking: Given the degree of manipulation k,
the strategy randomly shuffles the first k ranks. This strategy can be either applied
alone or in combination with the truncation strategy.

Truncation Including lower ranked alternatives in one’s preference ranking
might increase the chance of being matched to them, as it increases the potential
number of partners one can be matched with. By stating these lower ranked alternatives as unmatchable, a participant could end up being matched to a more
preferred alternative, thus increasing its benefit. However, stating otherwise acceptable alternatives as unmatchable also increases the chance of the participant
being unmatched in the final allocation, in case other participants have a higher
ranking in the rankings of the remaining alternatives. This consideration describes
the truncation strategy defined by Roth and Rothblum (1999). Given the true preference ranking of length n of user i, a truncation is defined as the preference ranking
that contains the first k users, k < n, in the same order as the true preferences. Roth
and Rothblum (1999) showed that truncation strategies dominate non-truncation
strategies under certain assumptions for the preference rankings, making them an
interesting candidate for the following evaluation.
Truncation strategies involve an inherent trade-off. Truncating to a high degree
aims to avoid being matched to less preferred alternatives, yet simultaneously
increases the probability of remaining unmatched. Roth and Rothblum (1999)
showed that for a given preference set, the number of participants benefiting from
truncation is small, yet its behavior under other algorithms or indifferences in preference rankings remain to be explored. Using analytical models, Ehlers (2008)
extends the analysis of Roth and Rothblum (1999) for priority-based and linear
programming mechanisms, and shows that under certain assumptions (symmetric
information) the same result about truncation preferences holds.

188

Incentive Compatibility and Dynamic Allocation

Manipulation Learning
The previous two strategies are examples of fixed strategies. The downside of evaluating such strategies is that to be able to evaluate their performance, they would
have to be compared to the theoretically best strategy with the highest gains for the
manipulating users. Finding such an optimal strategy is a combinatorial problem,
especially for preferences with indifferences or incompleteness. For example, for
a preference list of length n, there are n! different re-ordered lists. Hence, a proxy
is needed as benchmark. One potential proxy is the use of learning algorithms
which allow the manipulating participants to (try to) learn good strategies. With
such an approach, manipulating participants are more flexible and not restricted
to the provided or pre-specified strategies. It is also a more realistic representation of participants trying to actively game the matching mechanism by learning
and adapting their strategies. Two learning algorithms are used in the following
evaluation.

Evolutionary-based Learning The first is an evolutionary learning algorithm
based on a genetic algorithm. As mentioned earlier, it has been previously shown
that GA’s work well in large search spaces, and the number of potential strategies
for preference manipulation indeed is considerable. Furthermore, a similar learning algorithm has been successfully applied for other economic decision scenarios,
see e.g. Haas et al. (2013). As a GA, the learning algorithm is initialized with a
starting population of randomly created strategies, with an equal mix of truncation
and re-ordering strategies. Each strategy is represented by the adjusted preference
ranking. It uses standard mutation and crossover operators to evolve the initial
population and selects the best performing strategy after the last evolution cycle.
For crossover, a cycle-crossover operator is used to create new strategies. The mutation operator randomly applies either re-ordering or truncation, or adds users to
the preference ranking (reverse truncation). The latter option is included in the mutation operator to allow users to potentially retract from too severe truncation.

Probe and Adjust The second learning algorithm is a simple learning algorithm
similar to reinforcement learning, Probe and Adjust (PA). This type of algorithm
has been previously used in several studies, for example learning in an oligopoly
game (Kimbrough and Murphy, 2009) or learning in strategic contexts (Haas et al.,
2013). It is an example of an adaptive local search learning algorithm. Starting

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

189

from a given solution, the algorithm explores the neighborhood of the solution by
slightly adjusting the current solution. After several rounds of exploration, during which the fitness of the new potential solutions are probed, the PA selects the
best performing solution in the neighborhood (potentially the current solution) as
the new solution. This is repeated for a certain number of adjustment rounds, after which the currently best solution is returned. Applied to learning manipulation strategies, PA starts with a neighborhood of randomly created truncated and
re-ordered strategies, and selects the best performing one. Then, for each of the
subsequent rounds, the neighborhood of the currently best strategy is created by
truncating, re-ordering, or adding user rankings to the current preferences. The
pseudocode of this learning algorithm is described in Algorithm 6 in Appendix D.

6.1.2. Effects of Preference Manipulation on Matching Outcome
Preference manipulation can have several effects on the outcome of preferencebased matching:
• Potential gains and losses of manipulating users
• Effects on non-manipulating users
• Different solution quality under submitted and true preferences
Whereas the effects on the users, whether manipulating or not, are intuitive, the
third point is very interesting as well. If manipulating users do not submit their
true preferences, the resulting solution of the matching might, for example, not
be stable under the true preferences. These effects are studied in the following
evaluation.

Simulation Specifics
For the simulation-based evaluation, different scenarios are studied. For a given
problem size, the main parameters that are varied are the percentage of manipulating users, and the type or degree of manipulation. For example, for a problem size
of 20 users per side, the number of manipulating users is varied from 1 to 20. The
evaluation only considers manipulation on one side, and manipulation from both
requesting and providing users is not considered. The rationale behind this is that
for the standard DA, manipulation is only (potentially) beneficial for one side, and

190

Incentive Compatibility and Dynamic Allocation

in order to be able to compare the results between algorithms, only one-sided manipulation is considered here. The manipulation type can be either a fixed strategy
(in this section: truncation) or a strategy determined through a learning process.
For truncation strategies, the degree of manipulation defines how much the preferences are truncated. For example, in a 20x20 scenario with initially complete preferences, truncation of degree 0.5 means that the truncated preferences have length
10. Similar to the simulations in the previous chapter, 100 independent repetitions
are made for each scenario to ensure a certain robustness of the findings.
For the GA-learning parameters, the same parametrization as in the previous chapter is chosen. The population size is 50, the mutation rate 0.2, and the crossover
probability 0.6. In total, the learning was evolved for 20 rounds1 , and in every
round each solution of the population was evaluated twice to get an estimate of its
fitness. For PA-learning, two settings are studied to examine the effect of the available learning time on the outcome. In the first setting, each potential solution in
the neighborhood is evaluated twice, and 20 rounds of adjustments are simulated.
In the second setting, the solutions are evaluated four times, and 50 adjustment
rounds are studied.
The following results involve complete preferences with indifferences (maximum
group size of 2) for non-manipulating users. Obviously, in case of truncation strategies the problem transforms into a setting with incomplete preferences (with the
manipulating users submitting incomplete preference lists). Complete preferences
were chosen because users can only remain unmatched if other users manipulate, which makes the comparison with the baseline case easier where all users
are matched. In addition, small indifference groups potentially improve the benefits from manipulation, as it is easier to switch between the indifference groups.
Hence, this represents an optimistic setting in which benefits from manipulation
can potentially be higher. Starting with incomplete preferences is, of course, also
possible in general.

Manipulation Effects with Truncation Strategies Figure 6.1 shows the average
gain or loss of manipulating and non-manipulating users (measured in absolute
rank differences) for two different problem sizes if they apply truncation strategies
1 The

number of learning rounds is, for a GA, rather small. However, up to 1000 different manipulation strategies are considered in this setting, and the GA seem to converge to a certain
manipulation strategy after this number of rounds.

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

191

and the solution is calculated with GATA. Green colors indicate an average gain,
and yellow and red colors an average loss. Figure 6.1a shows that depending on
the number of manipulating users and the degree of truncation there are scenarios
where applying these strategies are, on average, beneficial for manipulating users.
However, the figure also shows two aspects to applying truncation strategies: on
one hand, the average gain is quite low, which makes the average benefit of manipulation doubtful. On the other hand, especially when the profiles are truncated to
a high degree, the expected outcome from truncation is an actual loss (measured in
the absolute difference of preference ranks) compared to submitting the true preferences. This is not surprising, as higher truncation leads to an increased likelihood
of being unmatched in the solution.
Similar results can be seen in Figure 6.1b, which considers the same setting for a
50x50 scenario. Interestingly, in this case truncating the preferences only to a small
degree is less beneficial than in the 20x20 case. As the absolute gain or loss is
small, however, this should not be overemphasized. The benefit of truncation for a
medium range of truncation seems to be consistent over the two studied problem
sizes.
Considering the effects on non-manipulating users, Figure 6.2 shows that in most
cases, manipulation of other users has detrimental effects on users that submit
their preferences truthfully. Especially when a large number of users are manipulating, the chances for them to remain unmatched increases, which also increases
the chance of a non-manipulating user to be unmatched. There are some cases in
which manipulation leads to an average gain, but the very small absolute amount
indicates that these gains occur as other solutions are calculated, which potentially
result in a better welfare for all users. This can happen as GATA is a heuristic and
potentially leads to different solutions once it is recalculated.
Additionally, the average gain in ranking for manipulating users has been compared for several matching algorithms. Tables 6.1 and 6.2 show the results for
different degrees of truncation and number of manipulating users, respectively.
Over all scenarios, i.e., all combinations of number of manipulating users and degree of manipulation, the absolute gains are smallest for GATA. This result is also
statistically significant using Wilcoxon signed-rank test at the level p = 0.001 for
the non-normally distributed data. Detailed results for the interplay of truncation
degree and number of manipulating users can be found in Tables D.1 and D.2 in
Appendix D.

192

Incentive Compatibility and Dynamic Allocation

(a) 20x20 scenario

(b) 50x50 scenario
Figure 6.1.: Effects of Truncation Strategies on the Manipulating Users, GATA

Solution Quality under Manipulation Considering the potential introduction of
instability in the solution, Table 6.3 considers this aspect for several problem sizes
and algorithms. The table shows that for small degrees of manipulation, i.e., if
manipulating users only truncate the last users in their preference ranking, the
resulting solution is still stable. However, as the degree of truncation increases,
the number of blocking pairs that are introduced when the true preferences are
considered increases considerably. This effect can be observed over all matching
algorithms, which indicates that it is an inherent property of user manipulation.

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

193

(a) 20x20 scenario

(b) 50x50 scenario
Figure 6.2.: Effects of Truncation Strategies on the Nonmanipulating Users, GATA

This result is interesting for practical reasons. Firstly, while the matching algorithms still calculate solutions which are stable under the submitted preferences,
they are potentially unstable under the true preferences if some users manipulate.
Hence, there might be opportunities for users to act on these unstable pairs, which
would break up the matching solution. Secondly, as Knuth (1997) discusses, even a
small number of unstable pairs can have substantial consequences on the market,
and allocations that contain blocking pairs can lead to a failure of the entire market. For example, it can lead to behavior where users prearrange matches before
the actual allocation takes place (see Sönmez (1999); Kagel and Roth (2000)).
An interesting extension of this observation is to examine which users are involved
in these unstable pairs. Theoretically, if unmatched users realize that they can find

194

Incentive Compatibility and Dynamic Allocation

Table 6.1.: Absolute Preference Gain for Truncation Strategies for Different Truncation Degrees, 20x20 Users
Truncation Degree

DA

RSMA

Király

Shift

GATA

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-5.12
-1.79
0.98
1.84
1.87
1.74
1.39
1.10
0.72
0.30

-3.89
-0.58
2.10
2.84
2.77
2.50
2.01
1.66
1.12
0.53

-4.49
-0.94
1.58
2.18
2.11
1.85
1.44
1.10
0.73
0.30

-4.91
-1.40
1.22
1.76
1.63
1.42
1.10
0.81
0.55
0.21

-6.13
-2.70
-0.06
0.56
0.57
0.44
0.26
0.14
0.03
0.00

Table 6.2.: Absolute Preference Gain for Truncation Strategies for Different Numbers of
Manipulating Users, 20x20 Users
Truncation Degree

DA

RSMA

Király

Shift

GATA

1
2
4
6
8
10
12
14
16
18
20

-1.04
-0.53
-0.23
-0.03
0.21
0.67
0.78
0.67
0.85
0.86
1.11

.18
.08
.40
.87
1.05
1.34
1.50
1.54
1.68
1.67
1.85

-.61
-.15
.01
.32
.54
.81
.98
.96
1.13
1.19
1.29

-.23
-.36
-.12
-.07
.07
.33
.42
.50
.60
.72
.78

-1.06
-1.26
-1.11
-0.91
-0.70
-0.57
-0.44
-0.53
-0.35
-0.38
-0.28

other users which would prefer to be matched with them, there could be an iterative process of users breaking up and forming new pairs. It is unclear, however, if
such an iterative process would lead to a stable match or not. Hence, it is hard to
predict what actual effects such unstable pairs which were introduced by manipulated preferences have on the market (Knuth, 1997).

6.1.3. Robustness of Matching Algorithms against Manipulation
Simulation Specifics
As before, the effects of manipulation are studied with varying parameters for the
number of manipulating users, and the type of manipulation. In this section, the
two different learning procedures (GA-Learning and Probe-and-Adjust-Learning)

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

195

Table 6.3.: Effects of Manipulation on Stability, Measured in Number of Blocking Pairs
Problem Size

Degree Truncation

DA

RSMA

Király

Shift

GATA

20x20

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

45.65
23.14
8.23
2.88
1.29
0.53
0.26
0.08
0.03
0.01

44.92
22.31
7.58
2.65
1.10
0.46
0.23
0.07
0.02
0.01

43.60
19.98
6.61
2.23
0.86
0.37
0.20
0.05
0.02
0.00

42.90
19.74
5.95
1.74
0.67
0.23
0.10
0.04
0.00
0.00

43.28
20.52
6.48
2.11
0.79
0.28
0.14
0.04
0.01
0.00

50x50

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

104.50
30.65
7.19
2.10
0.61
0.13
0.05
0.01
0.00
0.00

102.35
29.37
6.70
1.86
0.48
0.05
0.01
0.00
0.00
0.00

93.19
25.96
5.71
1.46
0.42
0.05
0.00
0.00
0.00
0.00

93.29
23.94
3.79
0.71
0.09
0.02
0.00
0.00
0.00
0.00

93.75
24.71
4.13
0.86
0.14
0.01
0.00
0.00
0.00
0.00

are used. For complexity reasons, only one or two manipulating users are considered. As before, for each scenario 100 independent repetitions are made to study
the robustness of the results. In addition to the previously considered algorithms,
GATA-MIXED (GATA with starting solutions from DA and approximation algorithms, see Section 5.4.4) is included in the following evaluation to study the dependency of the results on the GATA starting solutions.

Effectiveness of Manipulation
Previous results for DA suggest that the number of users who can actually benefit
from manipulation is small. Hence, this aspect is considered first. Table 6.4 shows
for a given market setting, how many times manipulating users can actually benefit from manipulation. The table can be separated into three scenarios. The first
scenario studies truncation strategies for different problem sizes and algorithms.
The second and third scenarios are a small number of manipulating users applying
GA- or PA-Learning, respectively.
The table shows that, averaged over the independent repetitions the probability of
successful manipulation, i.e. an actual gain in welfare for the manipulating users,
varies with the manipulation strategy. For both GA- and PA-learning, the actual

196

Incentive Compatibility and Dynamic Allocation

Table 6.4.: Manipulation Effectiveness
Scenario

Man.
Users

Strategy

DA

RSMA

Király

Shift

GATA

GATAMIXED

Probability of Successful Manipulation

20x20

50x50

1
2

GA
GA

0.540
0.560

0.540
0.578

0.560
0.584

0.440
0.460

0.436
0.448

.570
.485

1
2

PA
PA

0.540
0.555

0.620
0.610

0.550
0.605

0.480
0.470

0.401
0.448

.480
.555

1-20
1
2

Truncation
Truncation
Truncation

0.269
0.120
0.182

0.334
0.205
0.235

0.287
0.143
0.205

0.257
0.159
0.191

0.176
0.117
0.114

0.271
0.127
0.184

1-50
1
2

Truncation
Truncation
Truncation

0.335
0.142
0.172

0.387
0.194
0.216

0.342
0.146
0.185

0.314
0.149
0.175

0.210
0.130
0.138

0.337
0.146
0.164

1
2

GA
GA

5.23
3.94

6.22
4.51

5.28
3.92

5.84
3.61

2.82
2.12

5.07
3.49

1
2

PA
PA

4.62
3.09

6.08
4.17

5.09
3.68

4.56
3.10

2.75
1.71

5.04
2.94

1-20
1
2

Truncation
Truncation
Truncation

3.01
6.80
4.86

3.56
8.23
4.66

3.10
6.91
5.02

2.80
7.82
4.26

1.99
5.73
3.21

3.08
7.73
4.68

1-50
1
2

Truncation
Truncation
Truncation

6.92
17.74
12.07

7.84
18.72
12.05

6.93
16.89
12.32

5.78
14.07
9.82

3.82
8.74
6.76

6.71
16.86
9.77

Average Gain in Case of Successful Manipulation

20x20

50x50

manipulation is successful in 40%-62% of the cases, depending on the applied
matching algorithm. For truncation strategies, the success probabilities are considerably smaller (between 11% and 39%). In other words, this means that even for
users with the ability to learn and adjust their manipulation strategy, the chances
of actually benefiting from such manipulation are comparably small. Interestingly,
the success probability is smallest for GATA. A potential explanation for this result
is that GATA can find multiple solutions of the same or similar quality and randomly selects one of them. This additional random element could make it harder
for users to learn a successful manipulation strategy.
Another related question is the degree to which they can improve their position in
the cases where manipulation is indeed successful. The second part of Table 6.4
provides the results for the given scenarios. Interestingly, the potential improvements in case of truncation strategies are higher than for the learning strategies.

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

197

A likely explanation for this behavior is that, in the truncation cases, users might
gain considerably if they submit a highly truncated preference list. However, as the
table only shows actual improvements, not average gains, such heavily truncated
strategies might also lead to being unmatched in other cases (thus potentially leading to lower average gains). In contrast, when users have to learn a manipulation
strategy, they might not learn these extreme truncation strategies as they potentially involve a higher risk of being unmatched.
Considering the influence of the the small size of indifference groups that are used
in this scenario, it can be expected that the previous results are optimistic in the
sense that larger indifference groups could make it less likely to benefit from manipulation (as the chances for being matched within the same indifference groups
are higher). In total, it can be seen that users can gain several preference ranks if
their manipulation is successful, yet the probability of a successful manipulation is
comparably small. This indicates that even if there are cases in which (as predicted
by theory) manipulation is possible, the likelihood and incentives of users to actually manipulate are rather small. In particular, they might not outweigh the risk of
being worse off from manipulation.

Robustness Against Learning
Figure 6.3 shows the effects when the manipulating users apply a GA-Learning algorithm to learn successful manipulation strategies. There are several interesting
facts than can be observed. First, the actual number of benefiting users, as shown
in Figure 6.3b, is roughly similar for the studied allocation mechanisms, yet both
GATA and Shift seem to have a lower probability of successful manipulation. Averaged per user, the likelihood that a manipulating user is actually benefiting is
between 44-58%. Second, Figure 6.3a reveals considerable differences in the profitability of manipulation. Manipulation is less beneficial if GATA is used as an
allocation mechanism, compared to the other algorithms (statistically significant
at the level p = 0.001, using Wilcoxon signed-rank tests and accounting for multiple comparisons). One potential explanation is that due to the high number of
random elements, thus unpredictability, of the GATA mechanisms, it is harder to
achieve substantial gains for manipulating users in these settings. Interestingly,
GATA-MIXED performs similarly to the other algorithms and is more beneficial
for manipulating users. A potential explanation for this finding is that the solutions

198

Incentive Compatibility and Dynamic Allocation

found by GATA-MIXED are structurally similar to the solutions of the approximation algorithms such as Király, whereas the GATA solutions are less similar.2

AverageGainperManipulatingUser

3.5

20x20users,GAͲLearning

3
2.5
DA
2

RSMA
Shift

1.5

Király
1

GATA
GATAͲMIXED

0.5
0
1

2
NumberManipulatingUsers

(a) Gain in Ranking, GA-Learning

AverageNumberBenefittingUsers

0.7

20x20users,GAͲLearning

0.6
0.5
DA
0.4

RSMA
Shift

0.3

Király
0.2

GATA
GATAͲMIXED

0.1
0
1

2
NumberManipulatingUsers

(b) Number of Benefiting Users, GA-Learning
Figure 6.3.: Average Gain and Number Benefiting Users for GA-Learning, 20x20 users

In particular, manipulation gains are lower for GATA than for Király, which indicates that the former is more robust against manipulation as the potential gains are
lower. This affects the corresponding trade-off between potential gain and potential loss that a manipulating user faces, making it less appealing to try preference
manipulation.
2 In

other words, the final GATA-MIXED solutions can be similar to the solutions found by the
approximation algorithms as they are, on average, better than DA solutions for scenarios with
incomplete preferences, thus being more likely to be propagated through the evolution rounds
of GATA. As the standard GATA uses only DA starting solutions, the variety of different propagated solutions might be higher, thus decreasing predictability and the probability of successful
manipulation.

6.1. STRATEGIC MANIPULATION IN RESOURCE ALLOCATION

4

20x20users,PAͲLearning

4
3.5
3

DA

2.5

RSMA
Shift

2

Király

1.5

GATA
1

GATAͲMIXED

0.5

AverageGainperManipulatingUser

AverageGainperManipulatingUser

4.5

0

199

20x20users,PAͲLearning

3.5
3
DA

2.5

RSMA
2

Shift

1.5

Király

1

GATA
GATAͲMIXED

0.5
0

1

1

2
NumberManipulatingUsers

2
NumberManipulatingUsers

(a) Gain in Ranking, PA-Learning, 20 Ad- (b) Gain in Ranking, PA-Learning, 50 Adjustment Rounds
justment Rounds
0.7

20x20users,PAͲLearning

0.6
0.5
DA
0.4

RSMA
Shift

0.3

Király
0.2

GATA
GATAͲMIXED

0.1
0

AverageNumberBenefittingUsers

AverageNumberBenefittingUsers

0.7

20x20users,PAͲLearning

0.6
0.5
DA
0.4

RSMA
Shift

0.3

Király
0.2

GATA
GATAͲMIXED

0.1
0

1

2
NumberManipulatingUsers

1

2
NumberManipulatingUsers

(c) Number of Benefiting Users, PA- (d) Number of Benefiting Users, PALearning, 20 Adjustment Rounds
Learning, 50 Adjustment Rounds
Figure 6.4.: Average Gain and Number Benefiting Users for PA-Learning, 20x20 users

To compare the robustness of the previous findings, another learning procedure,
Probe and Adjust, is used as well. Figure 6.4 shows the results of this scenario.
Compared to the GA-learning case, PA-learning yields similar results. The number
of adjustment rounds does not have a significant impact on the number of benefiting users. Figures 6.4a and 6.4b show that the average gain from manipulation
is very similar. Furthermore, the results show that the average gain from manipulation is smallest for GATA. In fact, the average gain in the GATA case is statistically significantly smaller than the gains for Király and the other algorithms
(Wilcoxon signed-rank test, p = 0.001). GATA-MIXED performs similarly to the
other algorithms and worse than GATA, which again indicates that the solutions
found by GATA-MIXED seem to be more similar to the other approximation algorithms rather than GATA. The findings are particularly interesting as Király performed best among the approximation algorithms, and GATA-MIXED yielded the
best overall results (see Section 5.4.4). The findings indicate that there is no difference in the strategic properties of Király and GATA-MIXED, yet GATA is more

200

Incentive Compatibility and Dynamic Allocation

robust against manipulation (lower gains for manipulating users). This implies
an interesting trade-off between algorithm performance and strategic considerations.

6.1.4. Discussion
This section considered the effects of preference manipulation for the algorithms
presented in Chapter 5. There are several interesting insights that can be derived
from the results.
First, the results in Section 6.1.2 indicate that manipulation through preference
truncation can be beneficial for users, yet the average gains of successful manipulation are rather small (see Table 6.1). In many cases, manipulation has negative effects on non-manipulating users. Furthermore, instability can be introduced
through manipulation, indicated by the number of blocking pairs in Table 6.3. Due
to the unpredictable consequences of these blocking pairs, this can be a potentially
serious challenge for the applicability and sustainability of the allocation mechanism in a Social Cloud.
Second, the evaluation in Section 6.1.3 shows that the number of cases in which manipulation actually is successful is rather small (in most cases between 11 and 62%),
and that learning-based algorithms are more successful than pure truncation strategies. This indicates that, while manipulation is possible in such markets, finding
a successful manipulation strategy is a non-trivial problem. Comparing different
learning algorithms, for the studied scenarios the average gain from manipulation
is smaller for GATA than for the other algorithms, in particular the best performing
approximation algorithm (Király). This means that users can gain less from manipulating, thereby potentially lowering their incentive to pursue such manipulation.
Considering the combined results with respect to performance (see Section 5.4.4)
and strategic properties, GATA provides similar performance as the best performing approximation algorithm, and simultaneously is more robust against manipulation. If performance aspects are of higher importance, GATA-MIXED can be used
as it yields a superior solution quality while retaining similar strategic properties
as, e.g., Király.

6.2. DYNAMIC RESOURCE ALLOCATION IN A SOCIAL COMPUTE CLOUD

201

6.2. Dynamic Resource Allocation in a
Social Compute Cloud
Commonly, preference-based resource allocation mechanisms assume that all participating users in a market submit their preference profiles to the mechanism, and
a solution is calculated through a specific algorithm in a batch-like procedure. This
only applies if the allocation happens once, or at certain time intervals in which all
users are free and can be re-matched. Realistic scenarios, however, might be more
dynamic. For example, during two time intervals where the matching is calculated,
some users might enter or leave the market, thereby creating intermediate supply
and demand. Another example is if certain matched users agree to be matched
longer than a time period, in which case they might not be available in the next
matching calculation (i.e., they form an “unbreakable” pair).
In such scenarios, the question is how such intermediate demand is handled. For
the purpose of this case study, intermediate supply or demand is defined as newly arriving or freed resource offers or requests that occur when either new users arrive
or existing matched users leave, thereby creating additional resource offers or requests in the system. As mentioned before, breaking current pairs and re-matching
all users on the market might not be feasible for technical (e.g., if a calculation
on a machine cannot be interrupted) or complexity reasons (i.e., recalculating the
optimal allocation takes too long). The option to not consider intermediate demand leads to a potentially considerable amount of idle resources, depending on
the length of the time intervals. On the other hand, matching only the available
intermediate supply and demand might lead to the creation of blocking pairs, especially with respect to the currently matched pairs.
This dynamic allocation of resources in a two-sided matching setting is the focus
of this section, which is an extended version of (Caton et al., 2014). Building on the
description of the Social Compute Cloud prototype presented in Chapter 2.1.4, this
case study uses the prototype for the dynamic allocation scenario. The additional
heuristics that are used to match intermediate supply and demand are described in
Section 6.2.1. Section 6.2.2 evaluates these mechanisms that can be used to capture
intermediate supply and demand. A discussion of the findings in Section 6.2.3
concludes this case study.

202

Incentive Compatibility and Dynamic Allocation

6.2.1. Approaches to Capture Intermediate Supply and Demand
The matching algorithms discussed in related literature are usually assumed to be
batch jobs. In this case, allocations are computed after certain time intervals. For
example, economic studies of allocation mechanisms in Cloud Computing often assume that allocations are computed hourly, often referring to Amazon EC2 where
users buy resources based on hourly usage. In the case of a Social Compute Cloud,
one can say that the allocation is calculated every x hours, where x is the predetermined lease period of a compute vessel.
While this type of allocation computation yields good results for the supply and
demand given at the time of the computation, it is unclear what happens in the case
of new or changing supply and demand. For example, users can offer/request new
resources or retract offers/requests in between two calculation intervals. It is clear
that if allocations are only (re)computed at predetermined time intervals, resources
will be idle and requests will be left (or become) unsatisfied. Existing preferencebased matching literature does not consider such settings. Therefore the following
solutions for dynamic supply and demand are proposed.
Disregard The “worst” solution would be to disregard any new incoming supply
or demand until the next time the allocation is computed. In this case, new
supply/demand would be idle until the next batch allocation, even if there
were corresponding demand/supply.
Optimal The “optimal” solution would be the immediate rematching of the entire
supply and demand. In this case, no (new) supply would be idle if an allocation was available, and the resulting allocation would always be stable (given
the right algorithms are used). However, this places additional requirements
on the system. Firstly, computing resources (e.g. VM’s) would have to be
migratable at any given point in time, and, secondly, the runtime of the allocation mechanism has to be short. For the implementation of the Social
Compute Cloud prototype, this is currently unachievable, as the underlying
framework does not yet support migration. Hence, this approach should be
considered the best benchmark from a system perspective, i.e. with respect
to performance criteria such as stability, welfare, and fairness.
Random Given the intermediate offers and requests at a given time, random allocation randomly matches offers and requests with the constraint that the
matched users have to be mutually acceptable. As only one constraint is used

6.2. DYNAMIC RESOURCE ALLOCATION IN A SOCIAL COMPUTE CLOUD

203

in the calculation, this algorithm has a fast runtime, yet cannot guarantee specific properties of the resulting match (such as stability).
Greedy In greedy allocation, an incoming or freed provider or requester is
matched, if applicable, to a currently unmatched, acceptable requester or
provider of the highest possible rank. It is greedy in the sense that it tries
to find the best intermediate match for the incoming or freed user. Similar as
the random allocation, properties such as stability cannot be guaranteed.
As the Random and Greedy heuristics are applied each time a new offer or request
arrives, or a matched user leaves, they particularly have to have a fast runtime in
order to cope with potentially frequent changes in the system. Note that both Random and Greedy are likely to yield unstable solutions, i.e., the consumer-provider
pairs in the market at the end of the lease period would not be the pairs that a stable allocation algorithm would yield. However, if it is assumed that matched pairs
cannot be migrated in between two batch allocations, this does not, per se, affect
the practical stability in between batch computations.
A final approach is to check if there is a match that would yield a stable solution, yet
does not require other users to be reallocated. It can be argued that this approach
would be subsumed by the above approaches, as the probability of achieving a
stable solution is low, and in the absence of a stable match, another approach would
be applied.

6.2.2. Evaluating the Effects of Dynamic Allocation
To study the applicability of the previously mentioned approaches to match intermediate supply and demand in a Social Cloud, the following evaluation shows the
effects of the approaches on standard preference-based matching solution characteristics: the number of matched pairs, stability, welfare, and fairness.

Simulation Specifics
For the evaluation, a Social Cloud with 200 users (100 on each side) with incomplete
preferences and indifferences is simulated (with an average preference list length
of 50). To model dynamic supply and demand, real user resource availability distributions derived from donations and resource availability in SETI@home (Javadi

204

Incentive Compatibility and Dynamic Allocation

et al., 2009, 2011) are used in the evaluation. The data from SETI@home represents
statistical clusters of users, that can be used in the simulator to define both when a
resource will become available and for how long, as well as when users will request
resources.

Considerations for Stochastic Participation
To study stochastic supply and demand, the four approaches mentioned above
are simulated to study how they support the system with new supply/demand
in between two batch-allocation computations. Intuitively, immediate rematching
should yield the best solutions, whereas leaving resources idle should be worst.
Random and greedy should be somewhere in between.
In the simulation, each user is drawn an (un)availability distribution from the
SETI@home distribution, which determines when and how long they will be
(un)available. Only available users are taken into account for resource matching.
At time points 155, 265 and 410, the batch allocation algorithm is run for the current
supply/demand.
40

NumberofMatchedPairs

35
30
25
20
15
10
5
0
0

50

100

150

Optimal

200

250
300
EventNumber

Random

Greedy

350

400

450

500

Baseline

Figure 6.5.: Number of Matched Pairs for Intermediate Supply and Demand

Figure 6.5 shows that the number of matched users fluctuates over time, and users
arrive and leave from the allocation depending on their (un)availability pattern. It
can be seen that, most of the time, the “optimal” matching yields the highest number of matched pairs, and both random and greedy yield fewer matched pairs. This

6.2. DYNAMIC RESOURCE ALLOCATION IN A SOCIAL COMPUTE CLOUD

205

can be explained by the lack of choice that incoming users have: in random and
greedy, only the currently unmatched users are suitable for matching, whereas the
“optimal” rematching can consider all available users at that time. Figure 6.5 also
shows the baseline scenario in which allocation only happens at pre-determined
time intervals. In this case, new requests and offers are only considered at predetermined time intervals, and if matched users become unavailable, the corresponding
request/offer is freed without being automatically reallocated. Hence, the baseline
scenario depicts the worst case, quasi-static scenario where intermediate demand
is not considered. It can be easily seen in Figure 6.5 that not considering intermediate supply/demand can lead to a significant amount of unused, unallocated
resources.

NumberofUnstablePairs

12
10
8
6
4
2
0
0

50

100

150

Random

200

250
300
EventNumber
Greedy

350

400

450

500

Optimal

Figure 6.6.: Instability Effects for Intermediate Supply and Demand

Figures 6.6, 6.7a, and 6.7b show the results for stability, welfare, and fairness respectively. These figures show that the immediate rematching performs well with
respect to welfare (allocating users close to their highest preference) and fairness
(balancing the two market sides), and always yields stable outcomes. Furthermore,
the greedy strategy often provides better welfare than the random strategy (Figure 6.6), and is computationally as efficient. Whereas the runtime for GATA per
allocation is around 10 seconds, both random and greedy run almost instantly, i.e.,
take milliseconds to compute. Similar results are obtained for the number of unstable pairs, which are most often lower for the greedy strategy. Figure 6.7b shows
that both greedy and random strategy are more beneficial for consumers, indicated
by the lower scores, especially compared to the optimal matching. This finding,

206

Incentive Compatibility and Dynamic Allocation

AverageRankofMatchedUser

3
2.5
2
1.5
1
0.5
0
0

50

100

150

200

250
300
EventNumber

Random

Greedy

350

400

450

500

Optimal

(a) Welfare: Average Rank of Matched Users
1

DifferenceinAverageRanks

0.5
0
0

50

100

150

200

250

300

350

400

450

500

Ͳ0.5
Ͳ1
Ͳ1.5
Ͳ2

EventNumber
Random

Greedy

Optimal

(b) Fairness: Welfare Distribution. Negative scores represent
solutions more beneficial to consumers.
Figure 6.7.: Comparison of Matching Heuristics for Intermediate Supply and Demand

along with the fact that greedy can sometimes yield worse results for welfare as
well, is not surprising as it primarily aims to give new, incoming users their highest priority, without considering the preferences of other users.
Overall, the results suggest that approaches for the intermediate rematching of supply/demand are necessary, and that on average the greedy heuristic performs well
with respect to welfare, despite one side being favored in the matching. This is
especially interesting if immediate rematching is not technically feasible.

6.3. SUMMARY

207

6.2.3. Discussion
Section 6.2.2 shows convincing evidence that dynamic supply and demand between batch allocation times, caused by leaving and arriving users, has to be considered. While these results show that the solution quality tends to be very good
and even close to optimal it is clear that continuously running the algorithms might
not be feasible due to their computational overhead. Hence, fast heuristics are
needed that are able to deal with changing supply and demand, although these
heuristics usually lack the solution quality of the other algorithms. For small problem sizes, it might still be feasible to run algorithms such as the GATA in a continuous setting.
One potential strategy to improve both allocation quality and runtime would be
to compute an initial solution with a fast algorithm, e.g. DA, and then leverage
users’ provided computational power to improve solution quality. This would give
users an incentive to provide resources for a co-operative infrastructure (Haas et al.,
2013).
Another issue with the matching algorithms considered here is that they currently
support only one-to-one matchings, i.e., they do not yet support multi-unit allocations. In some settings, users might contribute or request multiple units of resources (e.g., several VMs to run a compute-intensive job). This is particularly interesting if intermediate demand can be captured by (potentially already matched)
users with remaining capacity. This extension is an area of future work.

6.3. Summary
Besides the pure performance characteristics of the algorithms, there are several
other interesting aspects that can be studied in the context of preference-based resource allocation mechanisms. This chapter presented two case studies, each of
which investigating and highlighting such an aspect. Section 6.1 presented the first
case study which considers the manipulation of preferences and the effects on the
allocation and the users. After introducing existing results and commonly studied manipulation strategies, the evaluation shows that users might benefit from
manipulation, but the probability of successful manipulation is rather small. In addition, both manipulating and non-manipulating users face the risk of being worse

208

Incentive Compatibility and Dynamic Allocation

off from manipulation. This highlights the essential trade-off that users have to
consider in this context: Weighing the chances of improving one’s outcome versus
potentially increasing the chances of not being allocated.
Considering the difference between matching algorithms, GATA is more robust
against manipulation as the average gains from manipulation are smaller compared to other algorithms, for example Király. This is interesting, as lower (average) gains also provide lesser incentives for preference manipulation in the first
place, especially considering the potential drawback of being unmatched if manipulation is not successful.
Section 6.2 considered aspects of dynamic allocation in case users arrive or leave
in between the calculation of allocations. This might be relevant if either the recalculation of the allocation is infeasible from a time-perspective, or if existing
matched pairs cannot be broken up due to technical reasons. The evaluation shows
that even very simple heuristic algorithms to match such intermediate supply or
demand only introduce a small number of unstable pairs.
The findings have several implications for resource sharing in Social Clouds. First,
the study on incentive compatibility shows that not only do heuristics prove to be
robust against manipulation, the potential benefits that users can achieve with such
preference manipulation is small and often outweighed by potential losses (not
being matched, or being matched to a less preferred partner). This is a promising
result as it is in line with the social philosophy of a Social Cloud: sharing should be
voluntary and strategic considerations should be of lesser importance. Second, the
dynamic allocation scenario shows that preference-based matching is also suitable
for Social Clouds with fluctuating user participation (and thus, resource offers and
requests). The type of heuristic used for resource allocation should depend on the
available time and technical considerations (e.g., the feasibility of VM migration).
As the effects of leaving new resources idle (i.e., not allocating them until the next
allocation interval) can be severe, the studied heuristics provide the means to use
provided resources in a Social Cloud in an efficient manner. For the participating
users, the matching of intermediate demand and supply implies that resources can
be allocated and used once available, thereby avoiding cases of not utilized, offered
resources.
These two case studies present a good starting point to investigate further topics
of interest. For example, in the context of preference manipulation, the question of
robust strategies is important. Given different scenarios and matching algorithms,

6.3. SUMMARY

209

a manipulation strategy can be considered robust if it consistently provides good
results (improvements to non-manipulation) in many scenarios. The existence of
such robust strategies is a topic for further research.
In summary, Chapters 5 and 6 showed how heuristics can be used to implement
preference-based resource allocation in Social Clouds. Their flexibility allows for
the adjustment to different scenarios and optimization goals, and users benefit
from a similar or improved performance in comparison to existing matching algorithms. They provide little incentives to manipulate preference rankings as the
respective probability to actually benefit from manipulation is small. Furthermore,
they are able to handle dynamic allocation scenarios when user participation (and
thus, resource offers and requests) is stochastic.
As this thesis focused on one-to-one allocations, the presented work can be extended in order to adapt the heuristics for many-to-one or many-to-many scenarios,
when multiple resources can be matched to one offer/request. This is particularly
interesting if a resource offer can satisfy several requests, or a request requires a
large amount of resources such that it cannot be covered with one offer (e.g., complex calculations requiring several VMs). Similar to this work, the comparison of
the adapted heuristics with existing mechanisms for many-to-one allocations, as
well as the consideration of incentive compatibility in such settings, are topics that
require further research.

Part IV.
Finale

Chapter 7.
Conclusion

T

HE types of resources that are shared and exchanged on online platforms is
as varied as the platforms themselves. The engineering of such platforms is
a complex task, and several challenges have to be addressed. This includes the
technical implementation, the design of a resource allocation mechanism, the need
to provide incentives for user participation, and the consideration of potential user
behavior and its effects on the platform. This thesis considered resource sharing
through the concept of Social Clouds, where the exchange of resources involves
non-monetary mechanisms on the basis of underlying social relationships. In such
a system, existing direct and indirect connections correspond to a certain level of
trust, which can have various effects on the sharing behavior of users (e.g., the
relevance of specific incentives, and the applicability of certain allocation mechanisms), and thus also affects the design of such a platform with its corresponding
challenges.
This thesis focused on two coordination challenges in the design of a Social Cloud:
user incentives for participation, and non-monetary allocation mechanisms based
on two-sided matching. As achieving a critical mass of actively participating users
is crucial for a sharing platform, the first part of the thesis focused on the understanding of relevant user incentives and the design of incentive schemes that take
into account the social setting of the platform. In the second part, the thesis analyzed two-sided matching heuristics for the non-monetary allocation of resources.
In particular, it focused on aspects of solution quality, strategic considerations, and
dynamic allocation.

213

214

Conclusion

Section 7.1 summarizes the contributions of the thesis. Section 7.2 critically discusses the assumptions and limitations of this work, and closes with an overview
of future work.

7.1. Contribution
The design of coordination mechanisms for Social Clouds, specifically the identification and study of participation incentives as well as the design of mechanisms
for resource allocation, was the focus of this thesis. Its contributions to the design of a social sharing platform are threefold: 1) the technical implementation of a
simulation tool as complementary methodology to study such systems; 2) the understanding of relevant incentives for participation as well as the engineering of incentive schemes; 3) the design of preference-based matching heuristics as a means
to allocate resources on a non-monetary basis. The three contribution aspects are
discussed in more detail in the subsequent sections.

7.1.1. Simulation-based Approach to Study Social Clouds
Social Clouds and other resource sharing concepts are complex systems. Different
types of users (with respect to resource endowment, sharing motivations, preferences, etc.) participate on such platforms and influence the overall system outcome
through their individual interactions with each other. Due to this complexity, the
study of Social Clouds presents methodological challenges. While analytical modeling and prediction of (dynamic) system effects might be possible only for small
systems, the (prototypical) implementation of the platform provides the opportunity to receive feedback from real users, their characteristics and behaviors. Similarly, certain effects such as influence of self-representation on the sharing behavior
can be studied through laboratory experiments. However, not all problems can be
addressed with the mentioned methodologies. For example, it can be necessary
to predict effects of rule changes on the system before they are implemented, or
stress-tests need to be run to study scalability behavior.
For these reasons, this thesis advocated the use of a simulation-based approach
as complementary methodology for Social Clouds and similar systems. The simulation tool described in Chapter 2.2 provides added value to the methodological

7.1. CONTRIBUTION

215

toolkit for the study of social sharing systems. It provides functionality to model
the relevant aspects of Social Clouds, ranging from users, underlying relationships
and networks, and a variety of allocation mechanisms, and is easily extendable to
capture new scenarios. The usefulness of the simulation tool and its applicability
for Social Cloud scenarios was shown throughout the thesis. Chapter 4 applied the
simulation tool to study effects of different incentive schemes on (heterogeneous)
user types and overall system performance. Chapters 5 and 6 used the tool to analyze the performance of different two-sided matching mechanisms for resource
allocation.
Due to its architecture and extensibility, the simulation tool is not only usable for
Social Cloud settings, but is also able to capture a variety of other resource sharing
scenarios in the future. However, as mentioned before, such a simulation-based
approach can be seen as complementary methodology and does not replace prototypes, experiments or theoretical analysis. As simulation results crucially depend
on the correct modeling of assumptions of input parameters, several steps such as
verification and validation of the simulation model as well as sensitivity analyses
of the results need to be pursued to ensure the correctness of simulation studies.

7.1.2. Understanding Incentives for Participation
To achieve a critical mass of participating users, a resource sharing platform such
as a Social Cloud needs to provide appropriate incentives to potential users. This
was the focus of research question 1:
R ESEARCH Q UESTION 1 ≺ PARTICIPATION AND C ONTRIBUTION I NCENTIVES 
What are relevant user participation incentives for Social Clouds and how can they be
leveraged in the design of tailored participation and contribution schemes?
The first step to address this research question is to analyze how users interact
with a Social Cloud in order to determine different participation stages. As the
relevance of certain incentives might change over the time of user participation
on the platform, research question 1.1 aimed at distinguishing these participation
stages and identifying relevant incentives for the different stages.
R ESEARCH Q UESTION 1.1 ≺ I NCENTIVE E NGINEERING  What are the stages of
participation and the corresponding relevant incentives that users exhibit in Social Clouds?

216

Conclusion

Three different stages of participation can be identified that are relevant for Social
Clouds: the registration on the platform, the active participation on the platform,
and the adherence to social norms and behavior (see Section 3.2). Through a web
survey, relevant incentives for the first two stages were identified (Section 3.3). For
user registration and participation, a direct invitation from friends, general altruistic traits, and the perceived benefits from joining were considered most important,
whereas the prospect of monetary compensation was considered least important.
In addition, the setting of the sharing scenario, i.e., whether sharing occurs in private settings between friends or in professional settings between colleagues, also
influenced the relative importance of certain incentives such as monetary compensation or fun from participation.
Having identified different stages of user participation and the need to potentially
emphasize different incentives in these stages, as a next step this knowledge can
be applied to design incentive schemes for active user participation on the platform. This thesis advocated the use of the previously mentioned simulation tool as
complementary methodology for the engineering of such incentive schemes. To exemplify the usefulness of the simulation-based approach, Chapter 4 presented two
case studies. The aim of the first case study was to study the effects of an incentive
scheme on different user types, as stated in research question 1.2.
R ESEARCH Q UESTION 1.2 ≺ I NCENTIVE S CHEME D ESIGN  How can a simulationbased approach be leveraged in the design of incentive schemes for participation?
The simulation tool was used to study a complex Social Cloud scenario where resources were shared between heterogeneous user groups. The aim of the case study
was to analyze the effects a trading constraint on the users and the system. As an
analytical approach was not feasible due to the number of considered users and
their heterogeneity, the (dynamic) effects of the trading constraint on different user
types were studied in Section 4.2. The simulation revealed interesting effects on
the different user types, which were affected differently to a substantial degree.
In particular, it showed that selfish users suffer most from the introduction of a
participation constraint, while altruistic and other user types generally are also affected to some degree. The case study also showed that the results are sensitive
with respect to the distribution of user types, which is an important input factor
that has to be considered in the design process of Social Clouds. Overall, the case
study showed that a simulation-based approach can be used to study and predict

7.1. CONTRIBUTION

217

dynamic effects in Social Clouds, and to find robust parameters for an incentive
scheme by analyzing different user type distributions.
The second case study in Section 4.3 proposed an economic model how infrastructure resources that are needed to host a Social Cloud can be co-operatively
provided by the users of the platform themselves. In such a scenario, the incentives to provide resources to the infrastructure determine the amount of platform
resources and thus the applicability of such an approach. Therefore, research question 1.3 studied different contribution schemes how users can provide resources to
the infrastructure, and examined how the feasibility of the approach is determined
by the characteristics of the user base.
R ESEARCH Q UESTION 1.3 ≺ C O - OPERATIVE I NFRASTRUCTURES  What are the
effects of different contribution schemes on co-operatively provided infrastructure resources
for Social Clouds?
The contribution schemes distinguished between required contribution (in which
case each user has to provide a certain amount of resources) and voluntary contribution (users choose their level of contribution according to their specific characteristics). The results for the different contribution schemes showed that schemes
with required contribution mostly lead to a sufficient number of contributed resources and guarantee a high level of platform availability and performance. In
contrast, the applicability of voluntary contribution schemes largely depends on
the characteristics of the underlying user base. For example, in systems with a
high number of selfish users, required contribution schemes are more promising.
The results also showed, however, that the utility for users is higher in case of voluntary contributions, as they themselves determine their optimal level of resource
contribution.
Addressing research question 1, the findings showed that three distinct stages for
user participation in a Social Cloud can be distinguished, and that non-monetary
incentives such as altruism, fun, and expected reciprocity are the most relevant
incentives in such a setting. Additionally, a simulation-based approach can be used
to design incentive schemes and study dynamic effects on the Social Cloud. The
presented results are, however, only a first step for the engineering of incentives for
actual Social Clouds. As shown in the case studies, the effectiveness of a particular
incentive scheme depends on the distribution of user types, hence it is necessary to

218

Conclusion

identify the actual distribution for a given setting. In addition, it remains a largely
open question how incentives can be tailored to individual users or user groups in
Social Clouds.

7.1.3. Heuristics for Preference-based Resource Allocation
The design of resource allocation mechanisms for Social Clouds was the second coordination challenge addressed by this thesis. Due to the underlying social connections between users, a preference-based, two-sided matching approach was considered as a means to allocate resources through a non-monetary mechanism.
In preference-based matching, users are split into two sides: users requesting resources, and users providing resources. Each user specifies a preference ranking
with whom of the other (market) side they want to be matched. Based on these
preference rankings, a two-sided matching algorithm then calculates a solution that
determines which users are matched. The thesis focused on one-to-one matches,
i.e., one provider is matched with one requester. The complexity of calculating a solution depends on both the structure of the submitted preferences and the desired
goals that the solution should satisfy. Preferences can be categorized in complete
and incomplete preferences, depending if all users of the other side are matched
or not, and having indifferences between certain users or not. The standard goal
in preference-based matching is to calculate a stable solution, which means that no
two users can both benefit by bilaterally breaking from the solution and forming
a new pair. Other commonly considered goals are welfare, measuring the rank
that the average user is matched with, and fairness, which considers if welfare is
distributed equally between the two sides.
In the general case where preferences can include either incomplete lists or indifferences, finding a solution that satisfies the standard goals (e.g., finding a stable solution with best welfare) is NP-hard for all but a small number of scenarios. Furthermore, for certain combinations of preference structures and goals, approximation
algorithms have been developed. However, the downside of current approximation algorithms is that they are specialized on a certain set of preference structures
and goals, and are not designed to handle other cases. Furthermore, there are scenarios for which no approximation algorithms exist. Hence, the focus of this part
of the thesis was following research question:

7.1. CONTRIBUTION

219

R ESEARCH Q UESTION 2 ≺ R ESOURCE A LLOCATION  Which types of algorithms
provide a good combination of performance, flexibility, and strategic properties for nonmonetary, preference-based resource allocation?
As this is a multifaceted question and the performance of the considered algorithms
can be studied from different points of view, several other research questions have
to be answered first. These questions each consider a specific aspect of algorithm
performance or properties.
Especially in light of social resource sharing platforms, the goals of the platform
can be diverse and change over time. For this reason, and due to the computational
complexity of the underlying problem, heuristic algorithms to calculate a solution
to the preference-based matching were proposed in Chapter 5. Two heuristics were
considered to calculate solutions for preference-based matching: a Genetic Algorithm (GA) and a Threshold Accepting (TA) algorithm, as well as the combination
of the two. For the applicability of these heuristics, particularly with respect to
the quality of their solutions, research question 2.1 considered the performance of
heuristics and existing algorithms in different settings.
R ESEARCH Q UESTION 2.1 ≺ P ERFORMANCE OF P REFERENCE - BASED
M ATCHING  What is the performance of heuristics for preference-based matching
compared to existing matching mechanisms?
For different preference structures (complete and incomplete preferences with indifferences), the heuristics were compared with existing algorithms for the specific
setting, focusing on standard metrics such as stability, welfare, fairness, and the
number of matched pairs. For complete preferences, depending on the type of solutions with which it is initialized, the GA is able to improve the solution quality
of the initial solutions, and in particular provides better solutions in comparison to
the standard algorithms in this case. Similar results were obtained for the combination of GA and TA. TA alone, however, is not a fitting heuristic for this scenario
as its solution quality is worse than GA (Section 5.4.3). For incomplete preferences,
the relative performance of the heuristics is even better. Both GA and TA are useful heuristics in this scenario, and their respective solution quality is similar to the
average quality of the best approximation algorithm. Furthermore, the combined
GA and TA algorithm with mixed initial solutions consistently yields significantly
better results than all other algorithms (Section 5.4.4). In addition to increasing

220

Conclusion

the number of matched pairs, the heuristics also yield solutions with considerably
better welfare and fairness properties.
Overall, the results showed that heuristics do not only provide flexibility to cope
with various preference and goal combinations, but also perform similarly or better than existing algorithms. Combined with a relatively short runtime, heuristics
provide an ideal combination of flexibility, runtime, and solution quality for the
calculation of preference-based matching allocations. For resource sharing in social contexts, both the platform and the users benefit from the application of the
proposed heuristics (through an increased number of shared resources, and being
matched to a more preferred partner, respectively).
Besides solution quality, strategic behavior of users is another important aspect in
preference-based resource matching. Matching mechanisms determine the allocation on the basis of the submitted preferences. However, if the mechanism is not
incentive compatible, users might not submit their true preferences. One of the
fundamental results in preference-based matching states that there can be no algorithm that always yields a stable solution and for which no user has incentives
to manipulate the submitted preferences to the algorithm (Roth, 1982). In other
words, for the algorithms considered in this thesis, at least some users theoretically
can benefit from preference manipulation. The existence of manipulation has been
studied for only a few of the considered algorithms, and the effect of manipulation
considering the heuristics and different scenarios was an open question. This was
the focus of research question 2.2.
R ESEARCH Q UESTION 2.2 ≺ I NCENTIVE C OMPATIBILITY  What are the effects of
preference manipulation on the manipulating users, non-manipulating users, and the solution quality?
Submitting manipulated preferences has complex effects on the matching solution.
It can be beneficial for the manipulating user by being matched to a more preferred partner, yet it is also possible that the manipulating user is matched to a less
preferred partner or even remains unmatched. The results of the corresponding
evaluation in Section 6.1.2 showed that preference manipulation can indeed lead
to being matched with better partners for the manipulating users, as predicted by
theory. However, the likelihood of a successful manipulation is relatively small
(between 40-44% for the combination of GA and TA), and severe manipulation can

7.1. CONTRIBUTION

221

lead to the user remaining unmatched. Furthermore, the results showed that preference manipulation can introduce instability in the solution. For example, a solution which is stable under the submitted (manipulated) preferences can be unstable
under the true preferences.
The robustness of the heuristics and best performing approximation algorithms
was considered as well. For two scenarios with different numbers of users, the
results indicate that potential (average) gains from manipulation are smallest for
the combination of GA and TA. This indicates that it is harder for manipulating
users to find a beneficial manipulation strategy if heuristic algorithms are used
(Section 6.1.3).
The third aspect of algorithm performance in preference-based matching was the
study of their applicability in dynamic allocation scenarios. In contrast to standard
preference-based matching which assumes batch-like calculations of the solution
at certain time intervals, in between these time intervals new users might enter
or leave the platform, thereby creating new (intermediate) supply and demand.
Research question 2.3 studied the effects of such dynamic supply and demand on
preference-based matching.
R ESEARCH Q UESTION 2.3 ≺ D YNAMIC A LLOCATIONS  What are options to allocate dynamic supply and demand, taking into account potential existing matches?
Section 6.2 showed that not acknowledging this intermediate supply and demand
might leave a considerable amount of resources idle. Depending on the problem
size, the runtime of the algorithms, or technical restrictions considering the migration of data or entire Virtual Machines, it might not be feasible to immediately
recalculate the entire solution with all users that are currently matched. Hence,
two additional heuristics were suggested to cope with such intermediate supply
and demand. The evaluation in Section 6.2.2 showed that these heuristics are able
to match the otherwise unutilized resources, yet introduce some level of instability
in the (overall) solution and have varying effects on welfare and fairness. If the allocation does not have to be calculated instantaneously, the recalculation with GA
and TA leads to the best results. Overall, however, the suggested heuristics for intermediate allocation are a valid option to match resources in dynamic scenarios.
Summarizing the results, the studied heuristics are superior to existing algorithms
considering their ability to adapt to different preference structures and goal com-

222

Conclusion

binations. For example, the heuristics can be easily adjusted to optimize combinations of goal metrics that existing algorithms do not or cannot consider. Their
performance for commonly considered metrics is equal or superior to existing algorithms, they offer good characteristics with respect to strategic behavior of users,
and their runtime allows the application in dynamic scenarios. Due to this, the
heuristics provide benefits for the resource allocation in Social Clouds. Depending on the setting, the optimization goals of the allocation can be flexibly adjusted
while the resulting solution quality of the allocation remains consistently high. In
particular, the improvements to existing (approximation) algorithms imply benefits for the system and the users (e.g., through better welfare and fairness properties
of a solution).
However, despite these positive results, there are some limitations to the presented
approach. Heuristics, in contrast to exact or approximation algorithms, do not
guarantee certain bounds for solution quality. Although the results in Chapter 5
indicate otherwise, there might be cases where the worst-case performance of the
heuristics is worse than the lower quality bounds of approximation algorithms. In
addition, this thesis considered one-to-one matching scenarios. The heuristics can
be easily adapted for many-to-one scenarios, yet their performance evaluation in
such settings requires further evaluation. Two-sided matching with complex preference structures such as multiple attributes also remains an open topic. These
issues are addressed in the next section on open research questions and future
work.

7.2. Future Work
This section critically discusses the assumptions and limitations of this work and
presents an outlook on future work.

7.2.1. User Participation Incentives
Participation Incentives: Feedback from Prototypes and Real Platforms
The identification of different participation stages in Section 3.2 is based on a conceptual model derived from literature research and a comparison with similar systems and models. The web-based survey on the relevance of certain incentive types

7.2. FUTURE WORK

223

for different users, as presented in Section 3.3, does not claim to be representative
of the actual user population, which limits the generalizability of the results. Both
steps are necessary in the design of participation incentives for Social Clouds, and
the results provided in Chapters 3 and 4 can be considered a first step towards a
more comprehensive coverage of this topic. In particular, the relevance of the incentives studied in the web survey in Chapter 3 have to be extended to capture
users of an actually implemented (prototype) platform, in order to derive insights
for the specific platform.
The results can be extended by using prototype implementations or actual platforms to gain feedback about user behavior, their participation incentives and other
user characteristics. This can help to both refine the model of participation stages
as well as to obtain useful insights about the importance of the studied incentive
types for certain user groups. The technical prototype implementation of a Social
Compute Cloud as presented in Section 2.1.4 provides an ideal starting point for
such research. As a next step, this prototype can be made accessible to users with
the goal to obtain feedback about the usage and the relevance of certain participation incentives.

Agreement Design for Social Clouds
The relevant motivations and incentives to exchange and share resource in social
settings, in particular on a non-monetary basis, are different from the primarily
monetary-based resource exchange systems such as the procurement of services
from service providers. In the latter case, the details of the exchange are often specified in Service Level Agreements (SLA) which determine the functional and nonfunctional properties of the given exchange. That is, SLAs specify the attributes of
the service that is exchanged and the relevant payments. Also, penalties in case of
unsuccessful service provisioning are defined.
In a more social context, the use of standard SLAs and monetary-based penalties
might be detrimental or have serious consequences on the existing relationships
between the providing and consuming users. Hence, a new form of socially-aware
SLAs has to be developed to cope with such situations, and still be able to specify certain properties of the exchanged resources and services. First steps in this
direction have been pursued by Michalk and Haas (2011), who discuss why social
aspects should be considered in the definition of SLAs in social contexts. Addi-

224

Conclusion

tionally, potential options and details how such agreements can be represented in
a Social Cloud are discussed by Thal (2013).

Sustainable Infrastructures for Social Clouds
The model for co-operative infrastructure provisioning as discussed in Section 4.3
only considers the trade-off between providing resources to the platform and keeping resources for other usage. In addition, the sensitivity analyses of the model
with respect to different user characteristics assumed the existence of certain types
of utility functions, which might not necessarily reflect the real user behavior.
An immediate extension, in particular in the context of Social Clouds, is the augmentation of the model to capture not only the contribution to the platform, but
potential resource sharing between users as well. In this case, users have the option to either reserve resources for own usage, donate them to the infrastructure, or
share them with other users on the platform. Users might also have different preferences and incentives with respect to these options. To create a necessarily realistic
model and evaluation of such a combined approach, a Social Cloud prototype with
said capabilities seems to be the best approach to evaluate actual user behavior and
its effects on the co-operative infrastructure model. A second extension is a feasibility study of having a sustainable user-contributed infrastructure for volunteer
computing. Usually, the infrastructure for BOINC-like projects is centralized, and
only the computing jobs are sent and retrieved from clients. However, it might also
be an viable option to host the project infrastructure itself on the client resources.

System Development and User Interface Design
This thesis concentrated on identifying participation incentives for Social Clouds,
as well as algorithms to match offers and requests. In the design of a Social Cloud,
there are further tasks that require attention. The concepts of perceived usefulness
and perceived ease of use, which are the two central constructs of the Technology
Acceptance Model (Venkatesh et al., 2003), are essential in the adoption of new
technology, and also new platforms. Hence, one of the necessary areas of design
is the development of good user interfaces. For example, Seuken et al. (2010) discuss user interface design in the context of P2P storage sharing. In the context of
Social Clouds, the influence of user interfaces on user behavior (e.g., with respect
to specifying sharing preferences) is of particular interest.

7.2. FUTURE WORK

225

7.2.2. Preference-based Resource Allocation
Multi-Attributive Two-Sided Matching
Preference-based matching is a suitable approach to allocate resources on a nonmonetary basis. The considered algorithms in this thesis focus on one-to-one
matching, where one requesting user is matched with one providing user. This
model assumes that resources are either homogeneous, or that a match is only allowed if the provided resource is able to fulfill the request (e.g., the provided VM
satisfies all requirements with respect to computational power, memory, etc.). This
simple representation might not be suitable to capture more complex scenarios.
For example, users providing a certain part of their computer as VM might be able
to split this VM into smaller instances and satisfy multiple smaller requests simultaneously. A direct extension of the one-to-one model, thus, would be to capture
such many-to-one matching algorithms.
In the case of heterogeneous resources, multiple attributes are often used to characterize the resources’ properties. For example, a VM can be described by the number
of cores, the provided memory and storage, its availability, etc. In such a scenario,
users might have different preferences for the attributes, i.e., weight the relative
importance of the attributes. If alternatives are ranked differently with respect to
the attributes, i.e., the ranking of alternatives is not the same for all attributes, the
question arises how a resource allocation can be found. One approach, in this case,
is to aggregate the preference rankings for the attributes into one ranking based
on the relative weight for the attributes. Another approach is to define necessary
requirements for the considered attributes that the alternatives have to fulfill, and
rank the alternatives according to the fulfillment of these requirements.
The allocation of seminar slots to students is another example where such a multiattributive approach can be useful. In such a scenario, students have a preference
for certain seminars they want to attend, and seminar leaders also have preferences which students they want to have in the seminar. Diebold et al. (2014) study
the application of stable matching algorithms on course allocation problems. Decentralized allocation, in this case, can be quite complex and lead to considerable
inefficiencies. For example, to increase their chances of getting a seminar slot students might apply for many seminars in parallel, and only accept the slots that
they like most, leading to potentially open slots in seminars where allocated students are not interested in the slot anymore. A centralized allocation mechanism
can help to alleviate this situation. For example, seminar leaders can state their

226

Conclusion

preferences for students (e.g., grade average, number of relevant lectures attended,
etc.), and students can rank the seminars according to their liking. The centralized mechanism can then find an allocation that matches students to seminar slots
such that the overall allocation efficiency (e.g., with respect to stability or welfare)
is increased compared to decentralized allocation. Such an approach can utilize
additional constraints, such as the guarantee that each student is allocated a minimum number of seminar slots. The analysis of stability and other performance
metrics is an interesting case study in such a scenario.

Weighted Preferences
The two-sided matching algorithms considered in this thesis focus on unweighted
preferences, i.e., preference rankings represent qualitative priority structures that
can be represented as rank order lists. For example, preference rank 1 denotes
the most preferred alternative, rank 2 the second most preferred alternative, etc.
While this is a common assumption in two-sided matching literature, a more quantitative preference representation can be useful in certain scenarios. For example,
Irving et al. (1987) propose weighted preference lists where, instead of preference
ranks, users are ranked according to a numerical score. Through such a representation, users are able to express preferences in more details. For example, a users
might consider the difference between the two most preferred alternatives as considerably higher than the difference between the two least preferred alternatives.
By providing scores instead of priorities, such a more complex representation is
achievable.
When weighted preferences are considered, several aspects need to be addressed.
Pini et al. (2011b) argue that the performance metrics should be adjusted to capture
the new preference representation. Considering algorithms for finding solutions to
the matching problem, the performance of the studied algorithms needs to be evaluated for such a setting. Although the heuristics can be easily adapted to capture
weighted preferences, their performance with respect to the optimal solution, or in
comparison with other algorithms, are an interesting topic for future research.

Robust Strategies for Two-Sided Matching
The aim of studying strategic preference manipulation in this thesis was to gain
insight in the potential effects of manipulation as well as the robustness of the

7.2. FUTURE WORK

227

considered mechanisms against manipulation. With this foundation, the following interesting question lends itself: What is a good strategy for manipulation in
two-sided preference-based matching markets? As seen in Chapter 6.1, theoretical results suggest that in certain scenarios truncation of preferences is superior to
both random reordering as well as submitting the true preferences. Yet, this does
not answer the question how much the preferences should be truncated, and how
such a strategy depends on the strategies of other, potentially also manipulating
participants. Considering the practical importance of such manipulation, it is also
of interest if there are robust strategies that work well in (many) different settings,
thus providing useful guidelines for participants and market designers alike.
The evaluation in Section 6.1 can be considered a first step in studying such robust
strategies. A potential research direction is to apply evolutionary computing to
determine successful manipulation strategies. Alternatively, a tournament in the
spirit of Axelrod’s strategy tournament for the Prisoner’s Dilemma game can be
held, where strategies are played against each other to determine the most robust
strategy (Axelrod and Hamilton, 1981; Axelrod, 1997).

Dynamic Preference Rankings and Feedback Integration
The preference-based matching algorithms presented in Chapters 5 and 6 take as
input the preference profiles of participating users. These profiles do not need to be
static, and thus can potentially change over time. For example, in a network with
providing and requesting users, the relative ranking of other users might be influenced by the previous interaction and sharing experiences as well as the feedback
about such past interactions from other users.
A potential way to rank users in the preferences is to use a trust network with local
or global trust values for the users. These trust networks incorporate the feedback
of other users in addition to own experiences about transactions with other users
(see e.g. Petri et al. (2012) for the use of trust networks in P2P clouds). If users did
not interact with certain other users before, it might be difficult to obtain a preference ranking due to the lack of knowledge about these other users. By using trust
networks and the corresponding trust values, users might be able to improve the
accuracy of their preference rankings over time. Two aspects are of particular interest in this case. On the one hand, the relative positions of users in preference
rankings can by dynamic and change based on their previous transactions as well
as user feedback. Hence, it is necessary to study the factors that influence this

228

Conclusion

dynamic ranking. On the other hand, the feedback itself might not always be reliable itself, or even be manipulated by maliciously behaving users. The effects of
different feedback types on the resulting solution, thus, is an important topic that
requires attention.

Preference Creation and Elicitation
Preference rankings submitted by users are the main input of the two-sided matching algorithms in this thesis. From the perspective of the users and the platform
designer, the creation and elicitation of these preference rankings is an interesting
aspect. In a Social Cloud, the data about the underlying social network and connections between users can be used in the creation of preference rankings. However,
as the representation of connections is often binary, the interpretation of a user’s
social ties for the purposes of allocation is not immediately clear. There is no single
unified methodology for the interpretation of social ties, and which to use is often context dependent. To create a preference ranking from the social ties, several
methods could be applied either separately or in combination with one another: 1)
ask users to rank their friends; 2) leverage methods from social network analysis to
identify features of social ties that can be used to (artificially) construct preferences;
and 3) use social network and interaction theories to construct a social sharing and
interaction model, and tune this model over time based upon observed interactions
within the social network platform and the Social Cloud. Each of these approaches
have their advantages and disadvantages, and there might be other methods which
are not listed above as well.
The use of user generated lists has the advantages that it is easy to implement, requires no special permissions (other than access to the list of friends), and should
be closest to capturing the true preferences of the user. However, given the recent
trends in social network usage, the average Facebook user currently has 190 friends
(Backstrom et al., 2012), this approach would not scale as more friends joined the
Social Cloud, as it cannot be expected that users rank large numbers of friends.
In contrast, the use of computational methods has the main advantage that these
approaches can be scaled as the Social Cloud grows. The challenge, however, is
in the identification of appropriate methods and indicators. These approaches also
require more data from the social network platform, and are thus more invasive
into the user’s private sphere, which cannot be understated. A simple example
that can be used in a preference-like manner are constructs like circles in Google+
or relationship lists in Facebook, as these are often created or at least curated by the

7.2. FUTURE WORK

229

user, and represent either specific (sub)groups in the social network and/or relationship types that are “similar” in some way. It is also possible to compose more
complicated methods of assessing social ties with the use of indicators to assess the
properties of a social tie. Overall, however, identifying the best implementation(s)
for the creation and elicitation of preferences remains an interesting and open challenge.

Part V.
Appendix

Appendix A.
Additional Material for Incentive
Survey
Survey Questions
1) Have you ever shared (provided and/or received) resources online? (Examples
for resources: files, programs, photos, lectures notes, (working) documents, sample
solutions, storage, ...)

◦ Yes

◦ No

2) Over which platforms and/or communication channels have you shared (providedand/or received) resources? (multiple answers allowed)

◦ Social network (private, e.g. Facebook)
◦ DropBox
◦ Google Drive
eMule,
◦ P2P-Tools (BitTorrent,
Limewire, ...)
◦ Other:

◦ Social network (professional, e.g.
Xing)
◦ Own server (FTP)
◦ Microsoft SkyDrive
◦ E-Mail

3) Which of the following resources have you already shared (provided and/or
received)online? (multiple answers allowed)

◦ Files
◦ Music
◦ Movies
◦ Sample Solutions
◦ Storage

◦ Programs
◦ Photos
◦ Lecture Notes
◦ (Working) Documents
◦ Other:

233

234

Additional Material for Incentive Survey

4) With whom have you shared (provided and/or received) resources so far?
(multiple answers allowed)

◦ Family
◦ Friends (online)
◦ Classmates
◦ Other:

◦ Friends (real life)
◦ Friends of Friends
◦ Colleagues

5) How often do you share (provide and/or receive) resources on average? (multiple answers allowed)

◦ More often than once per day
◦ Once per day
◦ A few times a week
◦ Once per week
◦ A few times a month
◦ Once per month
◦ frequently than once per month
◦ Dependent on certain events
◦ Please specify, if it is dependent on certain events:
6) What are (would be) your reasons for sharing (providing and/or receiving)
resources? (1 = disagree strongly, . . ., 7 = agree strongly)

◦ Direct request
◦ Own benefit
◦ Other compensation (e.g. favor)
◦ Other (please specify):

◦ Helpfulness
◦ Monetary compensation
◦ Prestige / Reputation

7) In addition to the resources, which you have already shared, are there any
other resources you would like to share if the appropriate technology was available? (Except for any illegal share activities)

◦ No
◦ Yes, the following:

◦ Don’t know

8) Are you interested in sharing (providing and/or receiving) resources in general? (Examples for resources could be files, programs, photos, lectures notes,
(working) documents, sample solutions, storage, ...)

◦ No
◦ Yes, the following:

◦ Don’t know

9) Which incentives would be important/crucial for you to register with that network? (In this case, registration does not imply active usage, only the general access to the network to, for example, see the offered resources. 1 = disagree strongly,
. . ., 7 = agree strongly)

◦ Request of closer friends
◦ Prestige / Reputation
◦ Helpfulness / favor
◦ Other (please specify)

◦ Monetary Compensation
◦ Curiosity / Fun
◦ Own benefit

235

Additional Material for Incentive Survey

10) Please now imagine that you are registered in such a sharing network with
friends. Which incentives would be important/crucial to be an active user, i.e.
to actively participate and share resources? (1 = disagree strongly, . . ., 7 = agree
strongly)

◦ Request of closer friends
◦ Prestige / Reputation
◦ Helpfulness / favor
◦ Other (please specify)

◦ Monetary Compensation
◦ Curiosity / Fun
◦ Own benefit

11) Please imagine that classmates, colleagues or acquaintances from a professional platform, e.g. Xing, participate in a closed sharing network. Which incentives would now be important/crucial for you to register to that network? (In
this case, registration does not imply active usage, only the general access to the
network to, for example, see the offered resources. 1 = disagree strongly, . . ., 7 =
agree strongly)

◦ Request of closer friends
◦ Prestige / Reputation
◦ Helpfulness / favor
◦ Other (please specify)

◦ Monetary Compensation
◦ Curiosity / Fun
◦ Own benefit

12) Please now imagine that you are already member of such a sharing network
with acquaintances, classmates or colleagues. Which incentives would be important/crucial to be an active user, i.e. to actively participate and share resources?
(1 = disagree strongly, . . ., 7 = agree strongly)

◦ Request of closer friends
◦ Prestige / Reputation
◦ Helpfulness / favor
◦ Other (please specify)

◦ Monetary Compensation
◦ Curiosity / Fun
◦ Own benefit

13) Imagine a scenario where it is possible to provide storage on your hard disk
to other people for storing their data (e.g. backups, documents, photos, etc.) and
access it online. You can be sure that there is sufficiently high security so that
only the owner of the data can create, read, update and delete their data and that
you are protected from viruses/malware and legal liability for the data stored on
your hard disk. To which groups would you provide your storage (independent
of compensation)? (multiple answers allowed)

◦ Nobody
◦ Closer Friends
◦ Classmates / Colleagues
◦ Other (please specify)

◦ Relatives
◦ Friends of Friends
◦ Everybody

236

Additional Material for Incentive Survey

14) Now you would like to store your own (personal) data (backups, documents,
photos, ...) on another’s hard disk. The data is encrypted, so that only you can
can create, read, update and delete your data. Which type of relationship must
exist between you and the provider of the storage? (multiple answers allowed)

◦ Relatives
◦ Closer Friends (online)
◦ Classmates / Colleagues

◦ Closer Friends (in real life)
◦ Friends of Friends
◦ Other (please specify)

15) In the previous scenario, "storage" was the resource shared between users.
Imagine now that you share (provide and/or receive) another resource (e.g. photos, lecture notes, (working) documents,...). Would the group of people with
whom you share resources change?

◦ No

◦ Yes:

16) Here are a number of personality traits that may or may not apply to you.
Please mark as appropriate to indicate the extent to which you agree or disagree
with that statement. You should rate the extent to which the pair of traits applies
to you, even if one characteristic applies more strongly than the other.
I see myself as... (1 = disagree strongly, . . ., 7 = agree strongly)

◦ Extroverted, enthusiastic
◦ Dependable, self-disciplined
◦ Open to new experiences, complex
◦ Sympathetic, warm
◦ Calm, emotionally stable

◦ Critical, quarrelsome
◦ Anxious, easily upset
◦ Reserved, quiet
◦ Disorganized, careless
◦ Conventional, uncreative

17) Please choose your age.

◦ < 20
◦ 26 - 30
◦ 41 - 50
◦ Prefer not to say

◦ 20 - 25
◦ 31 - 40
◦ > 50

18) Please choose your gender.

◦ Female

◦ Male

19) What is the highest degree or level of school you have completed?
If currently enrolled, mark the previous grade or highest degree received.

◦ Nursery-high school, no diploma
◦ Some college, no degree
◦ Master’s degree (for example: MA,
MS, MEng, MEd, MSW, MBA)
◦ Doctorate degree (for example: PhD,
EdD)
◦ Other (please specify):

◦ High school diploma
◦ Bachelor’s degree (for example:
BA, AB,BS)
◦ Professional degree (for example:
MD, DDS, DVM, LLB, JD)
◦ Post-doctoral education

237

Additional Material for Incentive Survey

20) Please choose your profession.

◦ Trainee
◦ Employee
◦ Prefer not to say

◦ Student
◦ Self-employed
◦ Other (please specify):

SurveyDesign
Introduction
Page
Yes

Q1:
Previously
Shared?
No

Q2: Platforms
Q8: General
Interest?

Yes

Q3: Resources
No
Q4: Users

Q9: Private Network
Registration
Q10: Private Network
Participtation

Q5: Frequency
Q16: TIPI

Q11: Professional
Network Registration

Q6: Motivations
Q17: Age
Q7: Other
Resources

Q12: Professional
Network Participation
Q18: Gender
Q13: Storage Sharing
Groups

Previous Sharing
Q19: Degree

Q14: Storage Sharing
Relationship
Q20: Profession
Q15: Other Resources
Personal and Demographic
Questions

Last Page

Figure A.1.: Incentive Survey Logic

Participation Incentives

238

Additional Material for Incentive Survey

Additional Survey Data
3ODWIRUPV8VHGIRU3UHYLRXV5HVRXUFH6KDULQJ



3HUFHQWDJH











HU

WK

H

ULY
'

H
ULY
\'
N
6N
RU
Z O
HW QD
1 LR
DO V
FL IHV
6R 3UR


2

3

N

RU
Z

HU

UY

6H

H
JO

RR

Q
Z

3

*

2

HW
1
DO
FL

6R

R[
S%
UR
'

O
DL

(0

DW

ULY

3
H

Figure A.2.: Platforms Used for Previous Online Resource Sharing
7\SHVRI3UHYLRXVO\6KDUHG5HVRXUFHV



3HUFHQWDJH









H

DJ

HU

RU

WK

2

6W

HV

RW

1

H

XU

LF

FW

XV

0

/H

HV

XU

V

FW

OH

)L

3L

V
LH
RY
0
V
DP
RJ
3U
H
SO QV
P WLR
6D ROX
6
 WV
N HQ
RU XP
: RF
'



Figure A.3.: Resource Types Previously Shared Online
6WRUDJH6KDULQJZLWK'LIIHUHQW*URXSV



3HUFHQWDJH











)DPLO\
5HODWLYHV

&ORVH &ODVVPDWHV 1RERG\

)ULHQGV
&ROOHDJXHV

)ULHQGVRI (YHU\ERG\
)ULHQGV

2WKHU

Figure A.4.: User Groups with which Storage would be Shared

239

Additional Material for Incentive Survey
5HTXLUHG5HODWLRQVKLS7\SHIRU6WRUDJH6KDULQJ



3HUFHQWDJH











)DPLO\
5HODWLYHV

)ULHQGVUHDO &ODVVPDWHV
OLIH
&ROOHDJXHV

)ULHQGV
RQOLQH

2WKHU

)ULHQGVRI
)ULHQGV

Figure A.5.: Required Relationships for Sharing Storage

Correlation Tables for Incentive Survey
Ext.

Agr.

Consc.

Emoti.

Open.

Ext.
Agr.
Consc.
Emoti.
Open.

1.000
-0.080
,190*
,216*
,373**

-0.080
1.000
0.023
0.031
0.152

,190*
0.023
1.000
,229*
0.104

,216*
0.031
,229*
1.000
0.114

,373**
0.152
0.104
0.114
1.000

Req.
Help.
Benefit
Monet.
Other
Rep.

-0.016
0.115
0.170
0.109
,223*
0.157

-0.077
0.071
0.000
-0.180
-0.092
0.066

,228*
,188*
0.045
0.000
-0.097
-0.023

0.159
0.055
0.043
0.044
0.165
-0.018

0.065
,199*
0.115
0.127
0.095
0.134

Table A.1.: Spearman-Rho Correlation Table for Incentives for Previous Sharing, Part 1

240

Additional Material for Incentive Survey

Req.

Help.

Benefit

Monet.

Other

Rep.

Ext.
Agr.
Consc.
Emoti.
Open.

-0.016
-0.077
,228*
0.159
0.065

0.115
0.071
,188*
0.055
,199*

0.170
0.000
0.045
0.043
0.115

0.109
-0.180
0.000
0.044
0.127

,223*
-0.092
-0.097
0.165
0.095

0.157
0.066
-0.023
-0.018
0.134

Req.
Help.
Benefit
Monet.
Other
Rep.

1.000
,301**
,224*
-0.044
0.086
-0.026

,301**
1.000
0.176
-0.027
0.100
,349**

,224*
0.176
1.000
0.163
,214*
0.072

-0.044
-0.027
0.163
1.000
,317**
,281**

0.086
0.100
,214*
,317**
1.000
,508**

-0.026
,349**
0.072
,281**
,508**
1.000

Table A.2.: Spearman-Rho Correlation Table for Incentives for Previous Sharing, Part 2

1.000
-0.080
.190*
.216*
.373**

0.069
.215*
.205*
.204*
0.126
.206*

.280**
.189*
.249**
.219*
-0.002
.199*

Ext.
Agr.
Consc.
Emoti.
Open.

Req.
Mon.
Rep.
Cur.
Help.
Ben.

Req.
Mon.
Rep.
Cur.
Help.
Ben.

Ext.

0.077
0.027
.220*
0.168
.312**
0.096

0.126
-0.091
0.089
0.123
.214*
0.020

-0.080
1.000
0.023
0.031
0.152

Agr.

0.080
0.022
-0.005
0.013
0.146
.193*

-0.040
-0.058
-0.043
-0.002
0.004
0.067

.216*
0.031
.229*
1.000
0.114

Emoti.

0.157
0.036
0.110
0.156
.257**
.191*

0.108
-0.064
0.063
0.102
0.133
0.137

.373**
0.152
0.104
0.114
1.000

Open.

.583**
0.075
.308**
.314**
0.088
.231*

1.000
0.101
.322**
.230*
.286**
0.122

0.069
0.126
0.096
-0.040
0.108

Req.

0.105
.772**
.367**
0.066
0.010
.222*

0.101
1.000
.371**
-0.019
.202*
0.162

.215*
-0.091
0.097
-0.058
-0.064

.204*
.290**
.731**
.409**
0.119
0.128

.322**
.371**
1.000
.388**
.350**
0.108

.205*
0.089
0.112
-0.043
0.063

.261**
-0.047
.380**
.660**
.263**
0.106

.230*
-0.019
.388**
1.000
.387**
0.166

.204*
0.123
0.074
-0.002
0.102

.204*
0.124
.386**
.348**
.520**
0.003

.286**
.202*
.350**
.387**
1.000
0.053

0.126
.214*
0.009
0.004
0.133

Correlations Private Network
Participation
Mon.
Rep.
Cur.
Help.

.231*
.205*
0.150
-0.054
0.052
.760**

0.122
0.162
0.108
0.166
0.053
1.000

.206*
0.020
0.015
0.067
0.137

Benefit

1.000
0.084
.279**
.272**
.243**
.375**

.583**
0.105
.204*
.261**
.204*
.231*

.280**
0.077
0.138
0.080
0.157

Req.

0.084
1.000
.378**
0.071
0.133
.253**

.279**
.378**
1.000
.418**
.188*
0.167

.308**
.367**
.731**
.380**
.386**
0.150

.249**
.220*
0.095
-0.005
0.110

.272**
0.071
.418**
1.000
.242**
0.103

.314**
0.066
.409**
.660**
.348**
-0.054

.219*
0.168
0.006
0.013
0.156

Active Sharing
Rep.
Cur.

0.075
.772**
.290**
-0.047
0.124
.205*

.189*
0.027
0.069
0.022
0.036

Mon.

Table A.3.: Spearman-Rho Correlation Table for Participation in Private Networks

0.138
0.069
0.095
0.006
0.143
0.018

0.096
0.097
0.112
0.074
0.009
0.015

.190*
0.023
1.000
.229*
0.104

TIPI
Consc.

.243**
0.133
.188*
.242**
1.000
0.137

0.088
0.010
0.119
.263**
.520**
0.052

-0.002
.312**
0.143
0.146
.257**

Help.

.375**
.253**
0.167
0.103
0.137
1.000

.231*
.222*
0.128
0.106
0.003
.760**

.199*
0.096
0.018
.193*
.191*

Benefit

Additional Material for Incentive Survey

241

Additional Material for Incentive Survey
242

Req.
Mon.
Rep.
Cur.
Help.
Ben.

Ext.
Agr.
Consc.
Emoti.
Open.

,320**
0.085
,200*
,351**
0.151
,287**

,207*
0.006
0.171
,328**
0.168
,259**

1.000
-0.080
,190*
,216*
,373**

Extr.

0.048
-0.124
0.137
0.118
,300**
0.007

-0.075
-0.178
0.123
0.126
0.142
-0.106

-0.080
1.000
0.023
0.031
0.152

Agr.

0.141
0.103
0.055
0.161
0.039
0.038

0.134
0.142
0.076
0.149
0.051
0.025

,190*
0.023
1.000
,229*
0.104

TIPI
Consc.

0.000
-0.061
-0.098
0.053
-0.030
0.099

0.023
-0.135
-0.085
0.138
-0.059
0.084

,216*
0.031
,229*
1.000
0.114

Em.
Stab.

0.161
-0.023
0.099
,182*
,235*
,232*

0.038
-0.151
0.069
,215*
0.180
0.145

,373**
0.152
0.104
0.114
1.000

Open.

,684**
,330**
,338**
,359**
,308**
,445**

1.000
,268**
,378**
,367**
,348**
,424**

,207*
-0.075
0.134
0.023
0.038

Req.

0.163
,807**
,333**
0.029
0.102
,305**

,268**
1.000
,451**
-0.058
,195*
,298**

0.006
-0.178
0.142
-0.135
-0.151

Monet.

,309**
,521**
,833**
,220*
,248**
,428**

,378**
,451**
1.000
0.173
,287**
,342**

0.171
0.123
0.076
-0.085
0.069

,440**
0.005
0.139
,779**
,350**
,234*

,367**
-0.058
0.173
1.000
,369**
,196*

,328**
0.126
0.149
0.138
,215*

,339**
0.171
,226*
,339**
,726**
,276**

,348**
,195*
,287**
,369**
1.000
,228*

0.168
0.142
0.051
-0.059
0.180

,453**
,363**
,291**
,193*
,268**
,842**

,424**
,298**
,342**
,196*
,228*
1.000

,259**
-0.106
0.025
0.084
0.145

Ben.

1.000
,284**
,290**
,351**
,356**
,487**

,684**
0.163
,309**
,440**
,339**
,453**

,320**
0.048
0.141
0.000
0.161

Req.

,284**
1.000
,429**
0.023
0.089
,438**

,330**
,807**
,521**
0.005
0.171
,363**

0.085
-0.124
0.103
-0.061
-0.023

Monet.

,290**
,429**
1.000
,280**
,278**
,355**

,338**
,333**
,833**
0.139
,226*
,291**

,200*
0.137
0.055
-0.098
0.099

,351**
0.023
,280**
1.000
,444**
,197*

,359**
0.029
,220*
,779**
,339**
,193*

,351**
0.118
0.161
0.053
,182*

Active Sharing
Rep.
Cur.

,356**
0.089
,278**
,444**
1.000
,336**

,308**
0.102
,248**
,350**
,726**
,268**

0.151
,300**
0.039
-0.030
,235*

Help.

,487**
,438**
,355**
,197*
,336**
1.000

,445**
,305**
,428**
,234*
,276**
,842**

,287**
0.007
0.038
0.099
,232*

Ben.

Correlations Professional Network
Participation
Rep.
Cur.
Help.

Req.
Mon.
Rep.
Cur.
Help.
Ben.

Table A.4.: Spearman-Rho Correlation Table for Participation in Professional Networks

Appendix B.
Extended Result Tables for
Co-operative Infrastructures
Table B.1.: Simulation Results for Enforced Fixed Contribution
x × ρ∗

Avg. Sys. Availability
Worst
Average
Case
Case

Avg. Contribution
Worst
Average
Case
Case

Ratio Prov. / Req.
Worst
Average
Case
Case

10

1
1.1
1.2
1.5

0.799
0.839
0.893
0.960

0.813
0.851
0.908
0.968

0.242
0.266
0.317
0.468

0.231
0.253
0.301
0.444

1.962
2.164
2.561
3.780

1.849
2.022
2.418
3.542

20

1
1.1
1.2
1.5

0.971
0.982
0.992
0.999

0.953
0.969
0.986
0.998

0.145
0.160
0.192
0.288

0.170
0.187
0.224
0.333

2.317
2.554
3.059
4.585

2.218
2.441
2.935
4.352

50

1
1.1
1.2
1.5

1.000
1.000
1.000
1.000

0.999
1.000
1.000
1.000

0.096
0.106
0.127
0.190

0.138
0.151
0.182
0.272

2.846
3.123
3.757
5.636

2.392
2.635
3.165
4.747

100

1
1.1
1.2
1.5

1.000
1.000
1.000
1.000

1.000
1.000
1.000
1.000

0.084
0.092
0.110
0.166

0.147
0.161
0.194
0.290

2.979
3.279
3.939
5.903

2.259
2.481
2.978
4.465

200

1
1.1
1.2
1.5

1.000
1.000
1.000
1.000

0.990
0.993
1.000
1.000

0.094
0.103
0.124
0.185

0.178
0.196
0.236
0.353

2.885
3.174
3.807
5.708

1.974
2.170
2.604
3.909

400

1
1.1
1.2
1.5

1.000
1.000
1.000
1.000

0.990
0.991
1.000
1.000

0.116
0.127
0.153
0.229

0.230
0.253
0.304
0.451

2.568
2.824
3.388
5.086

1.667
1.833
2.199
3.262

Number
Users

243

244

Extended Result Tables for Co-operative Infrastructures

Table B.2.: Simulation Results for Variable Fixed Contribution
ρ

Avg. Sys. Availability
Worst
Average
Case
Case

Avg. Contribution
Worst
Average
Case
Case

Ratio Prov. / Req.
Worst
Average
Case
Case

10

0.1
0.2
0.3
0.4
0.5

0.013
0.255
0.198
0.144
0.195

0.017
0.283
0.269
0.135
0.166

0.206
0.188
0.105
0.046
0.045

0.192
0.186
0.121
0.042
0.030

0.385
0.686
0.565
0.349
0.414

0.371
0.715
0.683
0.324
0.309

20

0.1
0.2
0.3
0.4
0.5

0.119
0.505
0.430
0.185
0.219

0.230
0.602
0.591
0.261
0.355

0.211
0.173
0.102
0.048
0.039

0.198
0.175
0.114
0.045
0.041

0.637
1.032
0.941
0.561
0.576

0.738
1.266
1.273
0.660
0.750

50

0.1
0.2
0.3
0.4
0.5

0.231
0.836
0.672
0.262
0.341

0.851
0.971
0.896
0.560
0.579

0.202
0.179
0.105
0.043
0.041

0.209
0.175
0.103
0.046
0.037

0.826
1.474
1.286
0.708
0.816

1.437
2.437
2.093
1.261
1.253

100

0.1
0.2
0.3
0.4
0.5

0.036
0.831
0.631
0.139
0.183

0.980
1.000
0.986
0.732
0.735

0.204
0.180
0.106
0.047
0.038

0.208
0.183
0.104
0.045
0.037

0.727
1.295
1.127
0.659
0.675

1.713
2.972
2.535
1.472
1.543

200

0.1
0.2
0.3
0.4
0.5

0.000
0.296
0.159
0.002
0.003

0.987
1.000
0.998
0.801
0.738

0.203
0.180
0.107
0.046
0.039

0.203
0.181
0.105
0.047
0.038

0.522
0.922
0.820
0.471
0.496

1.443
2.588
2.243
1.360
1.351

400

0.1
0.2
0.3
0.4
0.5

0.000
0.000
0.000
0.000
0.000

0.633
1.000
0.988
0.403
0.412

0.205
0.180
0.105
0.047
0.038

0.204
0.177
0.106
0.047
0.038

0.345
0.603
0.532
0.319
0.318

1.042
1.810
1.617
0.960
0.955

Number
Users

245

Extended Result Tables for Co-operative Infrastructures

Table B.3.: Simulation Results for Voluntary Variable Contribution, Baseline User Distribution
Number
Users

Relative
Price

Avg. Sys. Availability
Worst
Average
Case
Case

Avg. Contribution
Worst
Average
Case
Case

Ratio Prov. / Req.
Worst
Average
Case
Case

10

1.00
1.25
1.50
2.00
3.00

0.213
0.049
0.003
0.000
0.000

0.181
0.038
0.000
0.000
0.000

0.072
0.044
0.029
0.012
0.004

0.071
0.044
0.027
0.012
0.004

0.602
0.370
0.251
0.096
0.036

0.545
0.373
0.212
0.103
0.032

20

1.00
1.25
1.50
2.00
3.00

0.531
0.225
0.036
0.000
0.000

0.479
0.147
0.012
0.000
0.000

0.075
0.043
0.027
0.013
0.004

0.073
0.047
0.027
0.013
0.004

1.118
0.700
0.410
0.203
0.065

1.028
0.616
0.357
0.177
0.054

50

1.00
1.25
1.50
2.00
3.00

0.914
0.644
0.263
0.000
0.000

0.664
0.203
0.010
0.000
0.000

0.071
0.042
0.027
0.013
0.004

0.070
0.045
0.028
0.012
0.004

2.190
1.219
0.804
0.378
0.123

1.247
0.763
0.491
0.215
0.071

100

1.00
1.25
1.50
2.00
3.00

0.993
0.907
0.422
0.000
0.000

0.643
0.062
0.000
0.000
0.000

0.074
0.044
0.027
0.013
0.004

0.074
0.045
0.029
0.013
0.004

2.626
1.593
0.959
0.441
0.144

1.132
0.697
0.434
0.200
0.064

200

1.00
1.25
1.50
2.00
3.00

0.999
0.911
0.175
0.000
0.000

0.138
0.000
0.000
0.000
0.000

0.073
0.044
0.028
0.013
0.004

0.073
0.043
0.028
0.013
0.004

2.244
1.350
0.849
0.398
0.125

0.817
0.486
0.308
0.145
0.046

400

1.00
1.25
1.50
2.00
3.00

0.999
0.447
0.000
0.000
0.000

0.000
0.000
0.000
0.000
0.000

0.074
0.044
0.027
0.013
0.004

0.073
0.044
0.028
0.013
0.004

1.656
0.984
0.609
0.285
0.090

0.538
0.319
0.202
0.092
0.030

246

Extended Result Tables for Co-operative Infrastructures

Table B.4.: Simulation Results for Voluntary Variable Contribution, Selfish User Distribution
Number
Users

Relative
Price

Avg. Sys. Availability
Worst
Average
Case
Case

Avg. Contribution
Worst
Average
Case
Case

Ratio Prov. / Req.
Worst
Average
Case
Case

10

1.00
1.25
1.50
2.00
3.00

0.650
0.403
0.306
0.134
0.025

0.593
0.376
0.238
0.106
0.015

0.176
0.118
0.092
0.061
0.040

0.168
0.114
0.089
0.056
0.040

1.512
0.933
0.778
0.521
0.334

1.313
0.867
0.681
0.462
0.313

20

1.00
1.25
1.50
2.00
3.00

0.934
0.818
0.639
0.454
0.139

0.902
0.725
0.558
0.349
0.069

0.173
0.117
0.087
0.064
0.038

0.174
0.113
0.089
0.064
0.039

2.769
1.835
1.337
0.999
0.594

2.269
1.486
1.164
0.859
0.502

50

1.00
1.25
1.50
2.00
3.00

1.000
0.997
0.988
0.927
0.639

0.997
0.953
0.896
0.576
0.117

0.175
0.111
0.092
0.064
0.040

0.176
0.111
0.090
0.064
0.039

5.169
3.368
2.721
1.938
1.172

3.002
1.962
1.638
1.108
0.690

100

1.00
1.25
1.50
2.00
3.00

1.000
1.000
1.000
0.994
0.864

1.000
0.974
0.880
0.435
0.017

0.176
0.114
0.090
0.063
0.039

0.177
0.113
0.089
0.063
0.040

6.154
3.993
3.212
2.229
1.385

2.747
1.736
1.390
0.971
0.619

200

1.00
1.25
1.50
2.00
3.00

1.000
1.000
1.000
0.998
0.798

1.000
0.890
0.482
0.007
0.000

0.176
0.115
0.091
0.063
0.040

0.172
0.112
0.090
0.063
0.038

5.490
3.549
2.821
1.928
1.207

1.915
1.243
0.999
0.703
0.426

400

1.00
1.25
1.50
2.00
3.00

1.000
1.000
1.000
0.992
0.162

0.983
0.040
0.000
0.000
0.000

0.175
0.113
0.089
0.063
0.040

0.176
0.114
0.090
0.063
0.039

3.853
2.495
1.996
1.398
0.886

1.277
0.840
0.660
0.460
0.288

Appendix C.
Extended Results for
Preference-based Matching
Example for Preference Encoding and Genetic
Operators of the Genetic Algorithm
Example: Encoding, Crossover Operator, and Mutation Operator
Chromosome 1: (
i2 , j1  ,
i1 , j3  ,
i4 , j2  ,
i5 , j5  ,
i3 , ∅ ,
∅, j4 )
Chromosome 2: (
i2 , j3  ,
i1 , j1  ,
i4 , j5  ,
i5 , j2  ,
i3 , ∅ ,
∅, j4 )
After cycle crossover starting with the first gene:
New Chromosome 1: (
i2 , j1  ,
i1 , j3  ,
i4 , j5  ,
i5 , j2  ,
i3 , ∅ ,
∅, j4 )
New Chromosome 2: (
i2 , j3  ,
i1 , j1  ,
i4 , j2  ,
i5 , j5  ,
i3 , ∅ ,
∅, j4 )
Mutation operator, example for selecting two matched pairs and switching two
users:

i2 , j1  ,
i1 , j3  →
i2 , j3  ,
i1 , j1 
Mutation operator, example for selecting a mutation cycle:

∅, j4  ,
i1 , j3  ,
i3 , ∅ →
i1 , j4  ,
i3 , j3 

247

248

Extended Results for Preference-based Matching

Extended Result Tables for Preference-based Matching
Size

10x10

20x20

50x50

100x100

200x200

MGS

PM

DA (Best-Worst)

WO (Best-Worst)

GA-MIXED
(EW)

TA

GATADA

GATAMIXED

2

S
W
F

0.00 (0.00,0.00)
2.83 (2.74,2.91)
0.73 (0.67,0.80)

0.00 (0.00,0.00)
2.65 (2.58,2.74)
0.44 (0.41,0.53)

0.00 (0.00)
2.53 (2.61)
0.43 (0.16)

0.00
0.00
0.00

0.00
2.55
0.44

0.00
2.54
0.41

5

S
W
F

0.00 (0.00,0.00)
2.27 (1.97,2.61)
0.63 (0.42,0.92)

0.00 (0.00,0.00)
2.10 (1.79,2.52)
0.40 (0.29,0.52)

0.00 (0.00)
1.74 (1.82)
0.28 (0.07)

0.00
0.00
0.00

0.00
1.75
0.30

0.00
1.75
0.30

10

S
W
F

0.00 (0.00,0.00)
1.67 (1.22,2.27)
0.40 (0.12,0.88)

0.00 (0.00,0.00)
1.50 (1.17,2.06)
0.26 (0.10,0.45)

0.00 (0.00)
1.14 (1.16)
0.08 (0.04)

0.00
0.00
0.00

0.00
1.14
0.08

0.00
1.14
0.08

2

S
W
F

0.00 (0.00,0.00)
4.27 (4.15,4.40)
1.42 (1.24,1.61)

0.00 (0.00,0.00)
3.92 (3.82,4.01)
0.61 (0.60,0.60)

0.00 (0.00)
3.79 (3.88)
0.58 (0.21)

0.00
0.00
0.00

0.00
3.85
0.74

0.00
3.79
0.57

5

S
W
F

0.00 (0.00,0.00)
3.59 (3.11,4.23)
1.36 (0.78,2.28)

0.00 (0.00,0.00)
3.18 (2.77,3.69)
0.57 (0.38,0.74)

0.00 (0.00)
2.73 (2.84)
0.37 (0.07)

0.00
0.00
0.00

0.00
2.81
0.48

0.00
2.80
0.47

10

S
W
F

0.00 (0.00,0.00)
2.84 (2.08,3.82)
1.11 (0.36,2.33)

0.00 (0.00,0.00)
2.39 (1.89,3.00)
0.46 (0.33,0.69)

0.00 (0.00)
1.76 (1.86)
0.25 (0.04)

0.00
0.00
0.00

0.00
1.84
0.26

0.00
1.84
0.26

2

S
W
F

0.00 (0.00,0.00)
7.79 (7.35,8.19)
3.94 (3.11,4.69)

0.00 (0.00,0.00)
6.54 (6.40,6.69)
0.85 (0.77,0.94)

0.00 (0.00)
6.39 (6.48)
0.77 (0.37)

0.00
0.00
0.00

0.00
6.76
1.95

0.00
6.48
0.78

5

S
W
F

0.00 (0.00,0.00)
6.91 (5.76,8.52)
3.71 (1.59,6.32)

0.00 (0.00,0.00)
5.67 (5.23,6.18)
0.82 (0.59,1.12)

0.00 (0.00)
5.21 (5.34)
0.57 (0.06)

0.00
0.00
0.00

0.00
5.56
1.33

0.00
5.46
0.98

10

S
W
F

0.00 (0.00,0.00)
5.91 (4.50,8.06)
3.44 (1.23,6.47)

0.00 (0.00,0.00)
4.58 (4.01,5.22)
0.73 (0.53,1.12)

0.00 (0.00)
3.92 (4.07)
0.45 (0.05)

0.00
0.00
0.00

0.00
4.29
0.85

0.00
4.27
0.81

2

S
W
F

0.00 (0.00,0.00)
12.31 (11.26,13.44)
7.66 (5.95,9.39)

0.00 (0.00,0.00)
9.37 (9.18,9.57)
1.01 (0.88,1.05)

0.00 (0.00)
9.18 (9.29)
0.85 (0.27)

0.00
0.00
0.00

0.00
10.35
4.25

0.00
9.36
1.07

5

S
W
F

0.00 (0.00,0.00)
11.59 (9.35,14.56)
7.66 (3.76,11.89)

0.00 (0.00,0.00)
8.60 (8.13,9.10)
0.99 (0.73,1.32)

0.00 (0.00)
8.11 (8.27)
0.74 (0.08)

0.00
0.00
0.00

0.00
8.99
2.79

0.00
8.53
1.20

10

S
W
F

0.00 (0.00,0.00)
10.30 (7.69,13.99)
7.11 (2.55,12.04)

0.00 (0.00,0.00)
7.33 (6.75,7.95)
0.95 (0.59,1.30)

0.00 (0.00)
6.69 (6.86)
0.57 (0.07)

0.00
0.00
0.00

0.00
7.63
2.61

0.00
7.25
1.05

2

S
W
F

0.00 (0.00,0.00)
20.51 (17.70,23.70)
15.09 (11.01,19.33)

0.00 (0.00,0.00)
13.61 (13.34,13.87)
1.15 (0.96,1.15)

0.00 (0.00)
13.34 (13.48)
0.96 (0.18)

0.00
0.00
0.00

0.00
16.36
8.92

0.00
16.31
8.76

5

S
W
F

0.00 (0.00,0.00)
19.53 (15.17,24.99)
14.93 (8.24,21.88)

0.00 (0.00,0.00)
12.66 (12.16,13.17)
1.12 (0.89,1.50)

0.00 (0.00)
12.16 (12.33)
0.88 (0.11)

0.00
0.00
0.00

0.00
14.64
7.29

0.00
14.64
7.43

10

S
W
F

0.00 (0.00,0.00)
18.10 (13.27,24.10)
14.29 (6.90,21.72)

0.00 (0.00,0.00)
11.36 (10.78,11.97)
1.12 (1.01,1.45)

0.00 (0.00)
10.76 (10.95)
1.00 (0.11)

0.00
0.00
0.00

0.00
13.15
6.90

0.00
13.10
6.69

Table C.1.: Complete Preferences, Welfare Optimization

249

Extended Results for Preference-based Matching

Size

10x10

20x20

50x50

100x100

200x200

MGS

PM

DA (Best-Worst)

FE (Best-Worst)

GA-MIXED
(EW)

TA

GATADA

GATAMIXED

2

S
W
F

0.00 (0.00,0.00)
2.82 (2.78,2.83)
0.74 (0.59,0.92)

0.00 (0.00,0.00)
2.73 (2.75,2.70)
0.42 (0.27,0.57)

0.00 (0.00)
2.76 (2.61)
0.10 (0.16)

0.00
2.75
0.25

0.00
2.77
0.11

0.00
2.77
0.11

5

S
W
F

0.00 (0.00,0.00)
2.27 (2.17,2.30)
0.64 (0.19,1.14)

0.00 (0.00,0.00)
2.19 (2.23,2.42)
0.36 (0.02,1.02)

0.00 (0.00)
2.47 (1.82)
0.00 (0.07)

0.00
2.41
0.02

0.00
2.39
0.00

0.00
2.38
0.00

10

S
W
F

0.00 (0.00,0.00)
1.68 (1.53,1.33)
0.40 (0.03,0.33)

0.00 (0.00,0.00)
1.57 (1.54,1.94)
0.23 (0.00,0.76)

0.00 (0.00)
1.94 (1.16)
0.00 (0.04)

0.00
1.82
0.00

0.00
1.95
0.00

0.00
1.88
0.00

2

S
W
F

0.00 (0.00,0.00)
4.29 (4.19,4.40)
1.43 (1.13,1.91)

0.00 (0.00,0.00)
4.08 (4.06,4.12)
0.58 (0.29,0.92)

0.00 (0.00)
3.98 (3.88)
0.14 (0.21)

0.00
4.24
1.15

0.00
4.04
0.28

0.00
4.02
0.18

5

S
W
F

0.00 (0.00,0.00)
3.59 (3.29,4.16)
1.33 (0.41,2.79)

0.00 (0.00,0.00)
3.35 (3.29,3.85)
0.55 (0.01,2.01)

0.00 (0.00)
3.25 (2.84)
0.01 (0.07)

0.00
3.54
0.54

0.00
3.37
0.03

0.00
3.40
0.03

10

S
W
F

0.00 (0.00,0.00)
2.82 (2.49,3.01)
1.09 (0.09,1.98)

0.00 (0.00,0.00)
2.59 (2.50,3.11)
0.46 (0.00,1.69)

0.00 (0.00)
2.71 (1.86)
0.00 (0.04)

0.00
3.10
0.03

0.00
2.72
0.00

0.00
2.66
0.00

2

S
W
F

0.00 (0.00,0.00)
7.78 (7.38,7.93)
3.92 (3.03,4.33)

0.00 (0.00,0.00)
6.77 (6.71,6.91)
0.62 (0.12,1.40)

0.00 (0.00)
6.55 (6.48)
0.35 (0.37)

0.00
7.74
3.73

0.00
6.87
1.54

0.00
6.65
0.25

5

S
W
F

0.00 (0.00,0.00)
6.92 (5.97,8.93)
3.71 (1.13,6.95)

0.00 (0.00,0.00)
5.93 (5.81,6.91)
0.64 (0.01,3.84)

0.00 (0.00)
5.62 (5.34)
0.02 (0.06)

0.00
6.91
3.57

0.00
5.94
0.52

0.00
5.87
0.15

10

S
W
F

0.00 (0.00,0.00)
5.94 (4.77,9.31)
3.47 (0.60,8.26)

0.00 (0.00,0.00)
4.87 (4.77,5.96)
0.60 (0.01,3.69)

0.00 (0.00)
4.59 (4.07)
0.01 (0.05)

0.00
5.82
2.77

0.00
4.92
0.14

0.00
4.88
0.04

2

S
W
F

0.00 (0.00,0.00)
12.26 (11.26,12.68)
7.58 (5.78,8.34)

0.00 (0.00,0.00)
9.58 (9.54,9.88)
0.50 (0.08,1.81)

0.00 (0.00)
9.36 (9.29)
0.23 (0.27)

0.00
12.38
7.80

0.00
10.48
4.14

0.00
9.53
0.33

5

S
W
F

0.00 (0.00,0.00)
11.60 (9.36,15.59)
7.66 (3.30,12.96)

0.00 (0.00,0.00)
8.84 (8.71,9.97)
0.54 (0.01,4.60)

0.00 (0.00)
8.56 (8.27)
0.03 (0.08)

0.00
11.33
7.27

0.00
9.22
2.61

0.00
8.79
0.35

10

S
W
F

0.00 (0.00,0.00)
10.30 (7.77,20.30)
7.11 (2.11,19.20)

0.00 (0.00,0.00)
7.56 (7.49,8.86)
0.52 (0.01,4.72)

0.00 (0.00)
7.30 (6.86)
0.02 (0.07)

0.00
10.15
6.89

0.00
7.90
1.98

0.00
7.49
0.25

2

S
W
F

0.00 (0.00,0.00)
20.46 (17.77,20.12)
15.01 (11.00,14.56)

0.00 (0.00,0.00)
13.81 (13.77,13.93)
0.88 (0.18,1.67)

0.00 (0.00)
13.60 (13.48)
0.13 (0.18)

0.00
20.48
15.01

0.00
16.41
8.74

0.00
16.33
8.70

5

S
W
F

0.00 (0.00,0.00)
19.50 (15.14,29.27)
14.90 (8.10,26.51)

0.00 (0.00,0.00)
12.87 (12.83,13.10)
0.88 (0.06,2.11)

0.00 (0.00)
12.64 (12.33)
0.04 (0.11)

0.00
19.53
14.98

0.00
14.80
7.42

0.00
14.78
7.42

10

S
W
F

0.00 (0.00,0.00)
18.02 (13.35,43.24)
14.20 (6.91,42.15)

0.00 (0.00,0.00)
11.58 (11.51,11.90)
0.87 (0.06,2.70)

0.00 (0.00)
11.38 (10.95)
0.03 (0.11)

0.00
17.90
14.06

0.00
13.27
6.85

0.00
13.24
6.68

Table C.2.: Complete Preferences, Fairness Optimization

250

Size

10x10

20x20

50x50

100x100

Extended Results for Preference-based Matching

MGS

PM

DA (Best-Worst)

WO (Best-Worst)

GA-MIXED
(EW)

TA

GATADA

GATAMIXED

2

S
W
F

0.00 (0.00,0.00)
3.85 (3.74,3.95)
0.50 (0.44,0.60)

0.00 (0.00,0.00)
3.74 (3.67,3.81)
0.38 (0.36,0.38)

0.00 (0.00)
3.64 (3.70)
0.34 (0.12)

0.00
0.00
0.00

0.00
3.64
0.34

0.00
3.64
0.34

5

S
W
F

0.00 (0.00,0.00)
3.32 (3.05,3.68)
0.36 (0.22,0.62)

0.00 (0.00,0.00)
3.22 (3.00,3.60)
0.26 (0.18,0.36)

0.00 (0.00)
2.99 (3.05)
0.18 (0.06)

0.00
0.00
0.00

0.00
2.99
0.18

0.00
2.99
0.18

10

S
W
F

0.00 (0.00,0.00)
2.98 (2.71,3.42)
0.24 (0.06,0.50)

0.00 (0.00,0.00)
2.89 (2.68,3.29)
0.16 (0.06,0.27)

0.00 (0.00)
2.68 (2.69)
0.05 (0.02)

0.00
0.00
0.00

0.00
2.68
0.05

0.00
2.68
0.05

2

S
W
F

0.00 (0.00,0.00)
6.93 (6.79,7.08)
0.92 (0.73,1.09)

0.00 (0.00,0.00)
6.70 (6.62,6.80)
0.41 (0.36,0.46)

0.00 (0.00)
6.58 (6.67)
0.37 (0.08)

0.00
0.00
0.00

0.00
6.63
0.43

0.00
6.58
0.37

5

S
W
F

0.00 (0.00,0.00)
6.31 (5.89,6.84)
0.73 (0.37,1.30)

0.00 (0.00,0.00)
6.08 (5.79,6.49)
0.38 (0.28,0.48)

0.00 (0.00)
5.75 (5.85)
0.29 (0.03)

0.00
0.00
0.00

0.00
5.78
0.29

0.00
5.75
0.29

10

S
W
F

0.00 (0.00,0.00)
5.73 (5.22,6.43)
0.57 (0.20,1.33)

0.00 (0.00,0.00)
5.49 (5.13,5.96)
0.28 (0.16,0.42)

0.00 (0.00)
5.05 (5.12)
0.12 (0.03)

0.00
0.00
0.00

0.00
5.08
0.14

0.00
5.05
0.12

2

S
W
F

0.00 (0.00,0.00)
14.82 (14.43,15.25)
2.46 (1.79,3.20)

0.00 (0.00,0.00)
14.08 (13.92,14.25)
0.62 (0.57,0.64)

0.00 (0.00)
13.91 (14.02)
0.59 (0.05)

0.00
0.00
0.00

0.00
14.18
1.28

0.00
13.91
0.59

5

S
W
F

0.00 (0.00,0.00)
14.05 (13.26,15.18)
2.23 (0.95,4.04)

0.00 (0.00,0.00)
13.33 (12.97,13.75)
0.57 (0.44,0.76)

0.00 (0.00)
12.95 (13.08)
0.43 (0.03)

0.00
0.00
0.00

0.00
13.17
0.86

0.00
12.95
0.43

10

S
W
F

0.00 (0.00,0.00)
13.16 (12.23,14.53)
1.96 (0.69,3.89)

0.00 (0.00,0.00)
12.40 (11.97,12.89)
0.51 (0.36,0.70)

0.00 (0.00)
11.86 (12.00)
0.29 (0.02)

0.00
0.00
0.00

0.00
12.06
0.55

0.00
11.86
0.29

2

S
W
F

0.00 (0.00,0.00)
27.68 (26.72,28.83)
4.88 (3.17,6.70)

0.00 (0.00,0.00)
25.93 (25.71,26.18)
0.68 (0.56,0.77)

0.00 (0.00)
25.73 (25.85)
0.52 (0.06)

0.00
0.00
0.00

0.00
26.39
2.48

0.00
25.73
0.52

5

S
W
F

0.00 (0.00,0.00)
26.92 (25.49,28.88)
4.85 (2.25,7.75)

0.00 (0.00,0.00)
25.14 (24.76,25.56)
0.70 (0.56,0.79)

0.00 (0.00)
24.81 (24.96)
0.53 (0.04)

0.00
0.00
0.00

0.00
25.37
1.94

0.00
24.81
0.53

10

S
W
F

0.00 (0.00,0.00)
25.68 (24.12,27.81)
4.31 (1.52,7.23)

0.00 (0.00,0.00)
23.93 (23.44,24.46)
0.64 (0.51,0.85)

0.00 (0.00)
23.45 (23.64)
0.44 (0.02)

0.00
0.00
0.00

0.00
24.04
1.57

0.00
23.45
0.44

Table C.3.: Complete and Correlated Preferences, Welfare Optimization

251

Extended Results for Preference-based Matching

Size

10x10

20x20

50x50

100x100

MGS

PM

DA (Best-Worst)

FE (Best-Worst)

GA-MIXED
(EW)

TA

GATADA

GATAMIXED

2

S
W
F

0.00 (0.00,0.00)
3.83 (3.76,3.90)
0.50 (0.35,0.66)

0.00 (0.00,0.00)
3.81 (3.80,3.77)
0.28 (0.14,0.41)

0.00 (0.00)
3.84 (3.70)
0.07 (0.12)

0.00
3.83
0.08

0.00
3.82
0.08

0.00
3.84
0.07

5

S
W
F

0.00 (0.00,0.00)
3.34 (3.27,3.58)
0.37 (0.07,0.81)

0.00 (0.00,0.00)
3.28 (3.30,3.45)
0.23 (0.01,0.65)

0.00 (0.00)
3.48 (3.05)
0.00 (0.06)

0.00
3.48
0.01

0.00
3.45
0.00

0.00
3.48
0.00

10

S
W
F

0.00 (0.00,0.00)
2.97 (2.87,3.31)
0.23 (0.01,0.65)

0.00 (0.00,0.00)
2.93 (2.92,3.18)
0.15 (0.00,0.50)

0.00 (0.00)
3.12 (2.69)
0.00 (0.02)

0.00
3.06
0.00

0.00
3.10
0.00

0.00
3.12
0.00

2

S
W
F

0.00 (0.00,0.00)
6.93 (6.84,7.02)
0.90 (0.61,1.22)

0.00 (0.00,0.00)
6.82 (6.81,6.80)
0.33 (0.14,0.60)

0.00 (0.00)
6.82 (6.67)
0.04 (0.08)

0.00
6.86
0.17

0.00
6.82
0.11

0.00
6.82
0.04

5

S
W
F

0.00 (0.00,0.00)
6.31 (6.19,6.73)
0.73 (0.08,1.60)

0.00 (0.00,0.00)
6.20 (6.22,6.39)
0.31 (0.00,1.10)

0.00 (0.00)
6.24 (5.85)
0.00 (0.03)

0.00
6.37
0.01

0.00
6.22
0.00

0.00
6.24
0.00

10

S
W
F

0.00 (0.00,0.00)
5.73 (5.55,6.39)
0.58 (0.02,1.50)

0.00 (0.00,0.00)
5.60 (5.61,5.92)
0.25 (0.00,0.95)

0.00 (0.00)
5.84 (5.12)
0.00 (0.03)

0.00
5.96
0.00

0.00
5.71
0.00

0.00
5.84
0.00

2

S
W
F

0.00 (0.00,0.00)
14.80 (14.47,15.21)
2.43 (1.55,3.28)

0.00 (0.00,0.00)
14.27 (14.24,14.44)
0.29 (0.02,1.31)

0.00 (0.00)
14.21 (14.02)
0.02 (0.05)

0.00
14.70
1.94

0.00
14.28
0.83

0.00
14.21
0.02

5

S
W
F

0.00 (0.00,0.00)
14.05 (13.52,15.08)
2.23 (0.49,4.07)

0.00 (0.00,0.00)
13.53 (13.51,14.01)
0.30 (0.00,2.19)

0.00 (0.00)
13.48 (13.08)
0.00 (0.03)

0.00
14.07
1.04

0.00
13.54
0.13

0.00
13.48
0.00

10

S
W
F

0.00 (0.00,0.00)
13.17 (12.52,14.46)
1.99 (0.25,3.90)

0.00 (0.00,0.00)
12.61 (12.55,13.18)
0.29 (0.00,2.15)

0.00 (0.00)
12.62 (12.00)
0.00 (0.02)

0.00
13.30
0.13

0.00
12.76
0.00

0.00
12.62
0.00

2

S
W
F

0.00 (0.00,0.00)
27.67 (26.74,28.78)
4.87 (2.92,6.70)

0.00 (0.00,0.00)
26.19 (26.15,26.53)
0.29 (0.01,1.88)

0.00 (0.00)
26.08 (25.85)
0.02 (0.06)

0.00
27.49
4.51

0.00
26.51
2.24

0.00
26.08
0.02

5

S
W
F

0.00 (0.00,0.00)
26.91 (25.59,28.87)
4.85 (1.88,7.85)

0.00 (0.00,0.00)
25.42 (25.36,25.88)
0.28 (0.01,2.52)

0.00 (0.00)
25.34 (24.96)
0.01 (0.04)

0.00
26.81
4.38

0.00
25.54
1.49

0.00
25.34
0.01

10

S
W
F

0.00 (0.00,0.00)
25.65 (24.27,27.83)
4.26 (1.16,7.35)

0.00 (0.00,0.00)
24.22 (24.23,24.68)
0.26 (0.01,2.21)

0.00 (0.00)
24.12 (23.64)
0.00 (0.02)

0.00
25.55
2.88

0.00
24.42
0.78

0.00
24.12
0.00

Table C.4.: Complete and Correlated Preferences, Fairness Optimization

252

Size

10x10

20x20

50x50

100x100

Extended Results for Preference-based Matching

MGS

PD

Opt

DA

RSMA GSKirály
Modified

McShift
Dermid

GA

TA

GATA GATAMIXED

2

0.5
0.7
0.9
0.95

9.71
8.80
5.48
3.70

9.32
8.32
5.27
3.64

9.37
8.56
5.44
3.68

9.48
8.51
5.42
3.70

9.58
8.64
5.48
3.70

9.39
8.52
5.46
3.66

9.57
8.69
5.46
3.69

9.58
8.68
5.48
3.69

9.61
8.69
5.48
3.70

9.58
8.67
5.46
3.70

9.66
8.74
5.48
3.70

5

0.5
0.7
0.9
0.95

9.93
9.33
5.62
3.75

9.24
8.34
5.34
3.68

9.74
9.00
5.54
3.74

9.67
8.88
5.54
3.74

9.84
9.17
5.59
3.75

9.65
8.97
5.56
3.75

9.83
9.09
5.58
3.74

9.86
9.10
5.59
3.75

9.87
9.17
5.56
3.74

9.89
9.11
5.58
3.73

9.90
9.25
5.61
3.75

10

0.5
0.7
0.9
0.95

10.00
9.35
5.77
3.43

9.37
8.36
5.44
3.34

9.98
9.18
5.75
3.40

9.91
8.98
5.73
3.43

9.98
9.28
5.76
3.43

9.97
9.20
5.74
3.42

9.99
9.21
5.76
3.43

10.00
9.27
5.76
3.43

10.00
9.27
5.76
3.43

10.00
9.23
5.76
3.42

10.00
9.30
5.77
3.43

2

0.5
0.7
0.9
0.95

19.85
19.36
14.88
10.42

19.57
18.39
14.01
10.04

19.66
18.61
14.38
10.28

19.61
18.70
14.46
10.34

19.74
18.83
14.73
10.38

19.57
18.50
14.44
10.19

19.78
18.83
14.49
10.34

19.75
18.88
14.71
10.41

19.67
18.76
14.56
10.33

19.75
18.84
14.60
10.34

19.82
19.05
14.80
10.40

5

0.5
0.7
0.9
0.95

20.00
19.97
15.81
11.04

19.56
18.41
14.22
10.44

19.80
19.34
15.28
10.91

19.86
19.21
15.26
10.95

19.98
19.58
15.64
11.01

19.78
19.34
15.33
10.93

19.98
19.49
15.42
11.00

20.00
19.67
15.59
10.99

19.97
19.65
15.38
10.94

20.00
19.66
15.52
10.99

20.00
19.76
15.72
11.04

10

0.5
0.7
0.9
0.95

20.00
19.94
15.82
11.09

19.55
18.55
14.21
10.44

19.97
19.87
15.52
11.06

19.95
19.62
15.24
10.99

20.00
19.92
15.69
11.04

19.93
19.86
15.48
11.04

20.00
19.88
15.53
11.03

19.99
19.93
15.59
11.06

19.99
19.87
15.55
11.03

19.99
19.93
15.58
11.04

20.00
19.94
15.77
11.09

2

0.5
0.7
0.9
0.95

50.00
49.94
46.87
40.61

49.94
49.28
44.27
37.96

49.95
49.44
44.91
39.16

49.90
49.37
44.99
39.31

49.94
49.52
45.63
39.85

49.94
49.28
44.96
39.13

50.00
49.68
45.21
39.02

50.00
49.76
45.64
39.83

49.95
49.44
45.14
39.34

50.00
49.75
45.56
39.54

50.00
49.80
45.91
40.11

5

0.5
0.7
0.9
0.95

50.00
50.00
49.35
42.84

49.91
49.21
44.22
37.70

49.96
49.50
46.83
41.00

49.97
49.56
46.42
40.52

50.00
49.93
47.90
41.69

49.95
49.61
46.96
41.05

50.00
49.96
46.59
40.13

50.00
50.00
47.99
41.70

49.98
49.67
47.69
41.53

50.00
50.00
48.03
41.67

50.00
49.99
48.10
41.97

10

0.5
0.7
0.9
0.95

50.00
50.00
49.60
43.20

49.91
49.26
44.38
37.78

50.00
49.98
48.80
41.98

50.00
49.88
47.59
40.86

50.00
50.00
49.00
42.49

49.98
49.92
48.81
42.14

50.00
50.00
47.21
40.44

50.00
50.00
48.84
42.24

50.00
49.99
49.10
42.48

50.00
50.00
49.04
42.38

50.00
50.00
49.24
42.78

2

0.5
0.7
0.9
0.95

100.00
100.00
99.04
94.04

100.00
99.74
95.59
88.36

99.99
99.78
96.16
89.84

99.99
99.86
96.11
90.06

100.00
99.91
96.89
91.17

99.99
99.81
96.18
89.70

100.00
99.96
96.67
89.67

100.00
99.96
97.10
91.09

99.99
99.76
96.36
90.44

100.00
99.96
97.07
90.80

100.00
99.99
97.31
91.47

5

0.5
0.7
0.9
0.95

100.00
100.00
99.98
98.44

99.99
99.80
95.77
88.08

100.00
99.92
97.10
93.14

100.00
99.87
97.52
92.74

100.00
99.99
98.68
95.36

100.00
99.92
97.74
93.81

100.00
100.00
98.07
91.32

100.00
100.00
98.83
95.24

100.00
99.94
98.72
95.51

100.00
100.00
98.96
95.54

100.00
100.00
99.02
95.74

10

0.5
0.7
0.9
0.95

100.00
100.00
99.98
98.89

100.00
99.79
95.84
88.11

100.00
100.00
99.52
97.20

100.00
99.94
98.54
94.15

100.00
100.00
99.78
97.56

100.00
99.99
99.63
97.26

100.00
100.00
98.53
92.39

100.00
100.00
99.74
97.07

100.00
100.00
99.79
98.07

100.00
100.00
99.86
97.74

100.00
100.00
99.87
98.10

Table C.5.: Incomplete Preferences, 10x10 - 100x100 Users

253

Extended Results for Preference-based Matching

Size

200x200

500x500

MGS

PD

Opt

DA

RSMA

GSKirály
Modified

McShift
Dermid

GA

TA

GATA

GATAMIXED

2

0.5
0.7
0.9
0.95

-

200.00
199.98
197.88
191.40

200.00
199.98
198.06
192.60

200.00
199.98
198.07
192.43

200.00
199.99
198.50
193.87

200.00
199.99
198.23
192.50

200.00
200.00
198.59
192.78

200.00
200.00
199.07
194.09

200.00
199.98
197.99
192.92

200.00
200.00
199.07
193.93

200.00
200.00
199.17
194.27

5

0.5
0.7
0.9
0.95

-

200.00
199.99
197.99
191.42

200.00
199.99
198.73
194.13

200.00
199.99
198.55
194.89

200.00
199.99
199.52
197.85

200.00
199.98
198.76
195.40

200.00
200.00
199.46
194.67

200.00
200.00
199.67
197.44

200.00
199.99
199.26
197.56

200.00
200.00
199.71
197.65

200.00
200.00
199.86
197.94

10

0.5
0.7
0.9
0.95

-

200.00
199.98
197.88
191.33

200.00
200.00
199.72
199.00

200.00
199.99
199.16
196.60

200.00
200.00
199.87
199.50

200.00
199.99
199.80
199.00

200.00
200.00
199.87
195.35

200.00
200.00
199.90
199.25

200.00
199.99
199.95
199.63

200.00
200.00
199.93
199.54

200.00
200.00
199.95
199.65

2

0.5
0.7
0.9
0.95

-

500.00
500.00
499.68
496.81

500.00
500.00
499.78
497.28

500.00
500.00
499.66
496.90

500.00
500.00
499.75
497.84

500.00
500.00
499.75
497.13

500.00
500.00
499.94
497.72

500.00
500.00
499.99
498.67

500.00
500.00
499.69
497.06

500.00
500.00
499.99
498.67

500.00
500.00
499.99
498.73

5

0.5
0.7
0.9
0.95

-

500.00
500.00
499.69
496.73

500.00
500.00
499.88
498.32

500.00
500.00
499.73
497.50

500.00
500.00
499.97
499.32

500.00
500.00
499.89
497.97

500.00
500.00
500.00
498.70

500.00
500.00
500.00
499.14

500.00
500.00
499.80
498.36

500.00
500.00
500.00
499.09

500.00
500.00
500.00
499.45

10

0.5
0.7
0.9
0.95

-

500.00
500.00
499.72
496.80

500.00
500.00
499.94
498.87

500.00
500.00
499.83
498.39

500.00
500.00
500.00
499.94

500.00
500.00
499.94
499.39

500.00
500.00
500.00
499.24

500.00
500.00
500.00
499.94

500.00
500.00
499.94
499.86

500.00
500.00
500.00
499.89

500.00
500.00
500.00
499.95

Table C.6.: Incomplete Preferences, 200x200 - 500x500 Users

254

Size

10x50

10x100

20x50

20x100

Extended Results for Preference-based Matching

MGS

PD

Opt

DA

RSMA GSKirály
Modified

McShift
Dermid

GA

TA

GATA GATAMIXED

2

0.5
0.7
0.9
0.95

10.00
10.00
9.88
9.19

10.00
10.00
9.83
9.08

10.00
10.00
9.87
9.17

10.00
10.00
9.87
9.13

10.00
10.00
9.88
9.19

10.00
10.00
9.88
9.14

10.00
10.00
9.87
9.18

10.00
10.00
9.87
9.18

10.00
10.00
9.87
9.18

10.00
10.00
9.87
9.18

10.00
10.00
9.88
9.19

5

0.5
0.7
0.9
0.95

10.00
10.00
9.95
9.29

10.00
10.00
9.93
9.16

10.00
10.00
9.94
9.28

10.00
10.00
9.95
9.26

10.00
10.00
9.95
9.28

10.00
10.00
9.94
9.27

10.00
10.00
9.95
9.28

10.00
10.00
9.95
9.29

10.00
10.00
9.94
9.28

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.29

10

0.5
0.7
0.9
0.95

10.00
10.00
9.97
9.29

10.00
10.00
9.92
9.17

10.00
10.00
9.96
9.28

10.00
10.00
9.97
9.29

10.00
10.00
9.97
9.29

10.00
10.00
9.96
9.27

10.00
10.00
9.97
9.29

10.00
10.00
9.96
9.29

10.00
10.00
9.97
9.29

10.00
10.00
9.96
9.29

10.00
10.00
9.97
9.29

2

0.5
0.7
0.9
0.95

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.95

10.00
10.00
10.00
9.95

10.00
10.00
10.00
9.95

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.96

10.00
10.00
10.00
9.96

5

0.5
0.7
0.9
0.95

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.93

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10.00
10.00
10.00
9.94

10

0.5
0.7
0.9
0.95

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.90

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

10.00
10.00
10.00
9.91

2

0.5
0.7
0.9
0.95

20.00
20.00
19.87
17.78

20.00
20.00
19.71
17.31

20.00
20.00
19.83
17.60

20.00
20.00
19.80
17.61

20.00
20.00
19.83
17.76

20.00
20.00
19.86
17.59

20.00
20.00
19.83
17.71

20.00
20.00
19.87
17.75

20.00
20.00
19.86
17.64

20.00
20.00
19.87
17.75

20.00
20.00
19.87
17.78

5

0.5
0.7
0.9
0.95

20.00
20.00
19.87
18.23

20.00
20.00
19.64
17.56

20.00
20.00
19.85
18.14

20.00
20.00
19.86
18.05

20.00
20.00
19.86
18.19

20.00
20.00
19.83
18.09

20.00
20.00
19.87
18.17

20.00
20.00
19.87
18.21

20.00
20.00
19.87
18.16

20.00
20.00
19.87
18.21

20.00
20.00
19.87
18.22

10

0.5
0.7
0.9
0.95

20.00
20.00
19.84
18.08

20.00
20.00
19.59
17.34

20.00
20.00
19.82
18.01

20.00
20.00
19.84
17.93

20.00
20.00
19.84
18.06

20.00
20.00
19.83
18.00

20.00
20.00
19.84
18.04

20.00
20.00
19.84
18.07

20.00
20.00
19.83
18.03

20.00
20.00
19.84
18.07

20.00
20.00
19.84
18.08

2

0.5
0.7
0.9
0.95

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.85

20.00
20.00
20.00
19.86

20.00
20.00
20.00
19.87

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.85

20.00
20.00
20.00
19.87

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.86

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

5

0.5
0.7
0.9
0.95

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.81

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.87

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

20.00
20.00
20.00
19.88

10

0.5
0.7
0.9
0.95

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.87

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.90

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

20.00
20.00
20.00
19.91

Table C.7.: Incomplete Preferences, Asymmetric Sides, 10x50 - 20x100 Users

255

Extended Results for Preference-based Matching

Size

50x10

50x100

MGS

PD

Opt

DA

RSMA GSKirály
Modified

McShift
Dermid

GA

TA

GATA GATAMIXED

2

0.5
0.7
0.9
0.95

10.00
10.00
9.89
9.24

10.00
10.00
9.87
9.13

10.00
10.00
9.87
9.18

10.00
10.00
9.89
9.21

10.00
10.00
9.89
9.23

10.00
10.00
9.88
9.17

10.00
10.00
9.89
9.24

10.00
10.00
9.89
9.23

10.00
10.00
9.87
9.18

10.00
10.00
9.89
9.22

10.00
10.00
9.89
9.24

5

0.5
0.7
0.9
0.95

10.00
10.00
9.94
9.09

10.00
10.00
9.90
8.98

10.00
10.00
9.93
9.09

10.00
10.00
9.93
9.09

10.00
10.00
9.94
9.09

10.00
10.00
9.94
9.09

10.00
10.00
9.94
9.09

10.00
10.00
9.94
9.09

10.00
10.00
9.93
9.09

10.00
10.00
9.94
9.09

10.00
10.00
9.94
9.09

10

0.5
0.7
0.9
0.95

10.00
10.00
9.95
9.29

10.00
10.00
9.91
9.15

10.00
10.00
9.95
9.28

10.00
10.00
9.94
9.29

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.28

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.29

10.00
10.00
9.95
9.29

2

0.5
0.7
0.9
0.95

50.00
50.00
50.00
49.45

50.00
50.00
49.97
48.82

50.00
50.00
50.00
49.21

50.00
50.00
50.00
49.12

50.00
50.00
50.00
49.34

50.00
50.00
50.00
49.32

50.00
50.00
50.00
49.26

50.00
50.00
50.00
49.43

50.00
50.00
50.00
49.29

50.00
50.00
50.00
49.43

50.00
50.00
50.00
49.45

5

0.5
0.7
0.9
0.95

50.00
50.00
50.00
49.65

50.00
50.00
49.98
48.81

50.00
50.00
50.00
49.56

50.00
50.00
50.00
49.51

50.00
50.00
50.00
49.63

50.00
50.00
50.00
49.59

50.00
50.00
50.00
49.63

50.00
50.00
50.00
49.65

50.00
50.00
50.00
49.62

50.00
50.00
50.00
49.65

50.00
50.00
50.00
49.65

10

0.5
0.7
0.9
0.95

50.00
50.00
50.00
49.65

50.00
50.00
49.99
48.80

50.00
50.00
50.00
49.60

50.00
50.00
50.00
49.59

50.00
50.00
50.00
49.63

50.00
50.00
50.00
49.59

50.00
50.00
50.00
49.65

50.00
50.00
50.00
49.65

50.00
50.00
50.00
49.63

50.00
50.00
50.00
49.65

50.00
50.00
50.00
49.65

Table C.8.: Incomplete Preferences, Asymmetric Sides, 50x10 and 50x100 Users

256

Size

10x10

20x20

50x50

100x100

Extended Results for Preference-based Matching

MGS

PD

Opt

DA

RSMA GSKirály
Modified

McShift
Dermid

GA

TA

GATA GATAMIXED

2

0.5
0.7
0.9
0.95

9.76
8.76
5.41
3.58

9.23
8.24
5.29
3.55

9.41
8.38
5.33
3.56

9.44
8.52
5.37
3.58

9.52
8.60
5.41
3.58

9.43
8.43
5.37
3.55

9.61
8.59
5.41
3.58

9.61
8.63
5.41
3.58

9.63
8.65
5.40
3.58

9.57
8.64
5.39
3.58

9.70
8.74
5.41
3.58

5

0.5
0.7
0.9
0.95

9.99
9.08
5.62
3.65

9.29
8.19
5.35
3.55

9.79
8.81
5.58
3.60

9.71
8.68
5.57
3.64

9.93
8.90
5.61
3.65

9.79
8.75
5.50
3.63

9.94
8.90
5.62
3.64

9.93
8.93
5.61
3.65

9.93
8.98
5.62
3.65

9.95
8.91
5.58
3.65

9.97
9.03
5.62
3.65

10

0.5
0.7
0.9
0.95

9.97
9.18
5.74
3.44

9.28
8.32
5.47
3.39

9.85
8.93
5.66
3.42

9.77
8.78
5.67
3.44

9.94
9.06
5.72
3.44

9.85
8.96
5.65
3.43

9.97
9.12
5.72
3.44

9.95
9.05
5.73
3.44

9.95
9.10
5.73
3.44

9.93
9.04
5.73
3.44

9.95
9.15
5.73
3.44

2

0.5
0.7
0.9
0.95

19.96
19.34
15.00
11.12

19.48
18.33
14.29
10.68

19.51
18.55
14.63
10.90

19.58
18.54
14.72
10.94

19.72
18.75
14.89
11.06

19.60
18.52
14.58
10.95

19.82
18.84
14.82
11.04

19.77
18.90
14.88
11.09

19.67
18.78
14.75
10.98

19.75
18.81
14.85
10.98

19.87
19.00
14.96
11.11

5

0.5
0.7
0.9
0.95

20.00
19.87
15.65
10.83

19.42
18.30
14.36
10.29

19.80
19.29
15.06
10.65

19.84
19.17
15.03
10.68

19.95
19.58
15.36
10.80

19.78
19.14
15.17
10.67

19.97
19.45
15.29
10.74

19.98
19.67
15.39
10.82

19.94
19.53
15.15
10.74

19.98
19.65
15.28
10.77

20.00
19.77
15.54
10.83

10

0.5
0.7
0.9
0.95

20.00
19.92
15.61
10.92

19.42
18.33
14.06
10.38

20.00
19.61
14.99
10.71

19.95
19.39
15.04
10.75

19.99
19.68
15.41
10.87

19.99
19.56
15.10
10.69

20.00
19.70
15.14
10.83

20.00
19.78
15.33
10.90

20.00
19.72
15.16
10.83

20.00
19.76
15.28
10.86

20.00
19.87
15.52
10.90

2

0.5
0.7
0.9
0.95

50.00
49.93
46.85
39.62

49.83
49.01
43.94
37.43

49.91
49.25
44.67
38.26

49.85
49.14
44.66
38.54

49.89
49.41
45.24
38.99

49.84
49.15
44.62
38.29

49.98
49.53
44.98
38.43

49.99
49.60
45.48
39.03

49.89
49.27
44.87
38.36

49.99
49.57
45.34
38.80

49.99
49.69
45.70
39.24

5

0.5
0.7
0.9
0.95

50.00
50.00
48.87
41.56

49.81
48.96
43.80
37.49

49.92
49.54
46.43
39.69

49.92
49.46
46.03
39.65

49.96
49.84
47.37
40.58

49.94
49.48
46.58
39.50

50.00
49.92
46.04
39.45

50.00
49.96
47.22
40.63

49.93
49.73
47.13
40.05

50.00
49.95
47.27
40.36

50.00
49.98
47.65
40.94

10

0.5
0.7
0.9
0.95

50.00
50.00
49.19
41.63

49.79
48.99
43.89
37.17

50.00
49.94
47.87
39.74

49.98
49.83
46.53
39.60

50.00
50.00
48.14
40.74

49.98
49.96
47.95
40.06

50.00
50.00
46.58
39.47

50.00
50.00
48.11
40.70

50.00
49.99
48.13
40.19

50.00
50.00
48.22
40.58

50.00
50.00
48.54
41.14

2

0.5
0.7
0.9
0.95

100.00
99.99
99.15
93.26

99.95
99.61
94.88
87.08

99.95
99.60
95.40
88.55

99.96
99.61
95.71
88.70

99.99
99.71
96.29
89.93

99.94
99.68
95.52
88.46

99.99
99.95
96.10
88.46

99.99
99.96
96.66
90.10

99.94
99.59
95.57
88.98

99.99
99.96
96.51
89.74

100.00
99.97
96.87
90.47

5

0.5
0.7
0.9
0.95

100.00
100.00
99.94
97.71

99.98
99.60
94.72
87.08

99.98
99.85
96.98
92.10

99.99
99.77
97.02
91.93

100.00
99.97
98.51
94.35

100.00
99.83
97.42
92.72

100.00
100.00
97.06
90.35

100.00
100.00
98.57
93.86

99.97
99.89
98.31
93.94

100.00
100.00
98.61
93.91

100.00
100.00
98.74
94.75

10

0.5
0.7
0.9
0.95

100.00
100.00
99.99
98.26

99.97
99.58
94.90
87.06

100.00
99.97
99.41
95.02

99.99
99.89
98.16
92.72

100.00
99.99
99.68
95.92

100.00
99.98
99.43
94.93

100.00
100.00
97.92
91.05

100.00
100.00
99.64
95.38

100.00
100.00
99.71
95.85

100.00
100.00
99.69
95.62

100.00
100.00
99.82
96.41

Table C.9.: Incomplete and Correlated Preferences

Appendix D.
Extended Results for Preference
Manipulation
Pseudocode of Probe-and-Adjust Learning
Algorithm 6: Pseudocode of Probe and Adjust
Data: True Preference Profile
Result: Manipulated Preference Profile
1 begin
2
Create initial neighborhood from true preference profile;
3
for j ← 1to explorationRounds do
4
randomly select preferences from neighborhood;
5
submit these preferences to the matching algorithm, and save
outcome;
6
end
7
for i ← 1to steps do
8
create neighborhood based on current preference profile;
9
for j ← 1to explorationRounds do
10
randomly select preferences from neighborhood;
11
submit these preferences to the matching algorithm, and save
outcome;
12
end
13
select best performing preferences as new current profile;
14
end
15
return current best (manipulated) preference profile;
16 end

257

258

Extended Results for Preference Manipulation

Table D.1.: Absolute Preference Gain for Truncation Strategies, 1-10 Manipulating Users,
20x20 Users
Number Man.
Users

Truncation Degree

DA

RSMA

Király

Shift

GATA

1

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-8.83
-4.59
-0.23
0.43
0.57
0.75
0.49
0.51
0.37
0.14

-6.52
-2.68
1.22
2.05
2.06
1.80
1.33
1.21
0.90
0.47

-8.06
-3.43
0.53
1.15
1.02
0.91
0.72
0.52
0.38
0.14

-7.19
-2.77
1.20
1.38
1.36
1.29
1.00
0.65
0.58
0.18

-8.52
-4.10
-0.05
0.42
0.52
0.53
0.24
0.22
0.14
0.05

2

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-7.86
-4.78
0.23
1.53
1.55
1.52
1.09
0.81
0.44
0.23

-6.40
-3.31
1.12
2.20
2.20
1.92
1.30
1.10
0.49
0.21

-6.81
-3.44
0.86
1.84
1.66
1.59
1.19
0.84
0.50
0.26

-7.12
-3.65
0.89
1.57
1.40
1.25
0.88
0.67
0.35
0.13

-8.40
-4.92
-0.51
0.32
0.28
0.35
0.18
0.16
0.00
-0.04

4

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-6.98
-2.95
0.51
1.60
1.58
1.47
1.08
0.83
0.43
0.12

-6.03
-2.02
1.65
2.38
2.33
2.05
1.47
1.31
0.60
0.31

-6.13
-1.94
1.21
1.73
1.67
1.48
0.98
0.75
0.30
0.06

-6.30
-2.18
0.86
1.58
1.40
1.31
0.94
0.66
0.41
0.18

-7.68
-3.77
-0.57
0.23
0.23
0.34
0.17
0.14
-0.13
-0.05

6

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-6.05
-2.21
0.61
1.66
1.63
1.50
1.04
0.81
0.53
0.22

-4.85
-0.75
1.89
2.84
2.63
2.35
1.79
1.44
0.89
0.45

-4.94
-1.27
1.33
1.97
1.89
1.59
1.18
0.82
0.45
0.21

-5.66
-1.64
0.82
1.52
1.37
1.15
0.80
0.57
0.33
0.06

-6.83
-3.13
-0.23
0.29
0.42
0.34
0.14
0.01
-0.06
-0.05

8

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-5.29
-1.74
0.86
1.80
1.82
1.59
1.27
1.01
0.61
0.16

-3.98
-0.44
2.16
2.85
2.73
2.45
1.89
1.49
0.91
0.45

-4.62
-0.94
1.64
2.15
2.10
1.75
1.35
1.03
0.67
0.25

-5.21
-1.54
0.99
1.55
1.44
1.23
0.93
0.70
0.45
0.12

-6.27
-2.70
-0.16
0.62
0.51
0.36
0.23
0.12
0.16
0.13

10

0.95
0.9
0.8
0.7
0.6
0.5

-4.30
-1.03
1.22
2.13
2.15
1.98

-3.25
-0.04
2.21
3.02
2.94
2.64

-3.68
-0.51
1.68
2.39
2.30
2.00

-4.37
-1.26
1.23
1.92
1.73
1.40

-5.46
-2.43
-0.06
0.62
0.67
0.45

259

Extended Results for Preference Manipulation

Table D.2.: Absolute Preference Gain for Truncation Strategies, 10-20 Manipulating Users,
20x20 Users
Number Man.
Users

Truncation Degree

DA

RSMA

Király

Shift

GATA

10

0.4
0.3
0.2
0.1

1.76
1.38
0.91
0.47

2.18
1.85
1.30
0.49

1.60
1.22
0.84
0.31

1.07
0.85
0.57
0.19

0.24
0.09
0.11
0.10

12

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-3.82
-0.85
1.46
2.26
2.26
2.04
1.69
1.41
0.97
0.35

-2.87
0.23
2.42
3.20
3.10
2.73
2.30
1.94
1.39
0.58

-3.28
-0.17
1.97
2.61
2.52
2.06
1.65
1.26
0.89
0.27

-4.06
-0.93
1.35
1.89
1.75
1.46
1.13
0.87
0.59
0.19

-5.17
-1.99
0.01
0.84
0.78
0.44
0.35
0.22
0.06
0.04

14

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-3.96
-0.74
1.29
2.08
2.12
1.91
1.54
1.21
0.91
0.34

-2.72
0.44
2.41
3.08
3.01
2.74
2.30
1.92
1.48
0.68

-3.35
0.04
1.90
2.43
2.42
2.08
1.64
1.25
0.91
0.34

-3.96
-0.62
1.35
1.96
1.77
1.49
1.20
0.90
0.66
0.26

-5.17
-2.02
0.04
0.59
0.67
0.38
0.22
0.01
0.00
-0.04

16

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-3.29
-0.49
1.51
2.23
2.25
2.08
1.68
1.32
0.87
0.36

-2.21
0.57
2.60
3.14
3.17
2.92
2.46
2.01
1.46
0.68

-2.90
0.23
2.00
2.56
2.52
2.19
1.73
1.41
1.03
0.54

-3.49
-0.47
1.49
1.94
1.85
1.60
1.26
0.92
0.60
0.30

-4.83
-1.70
0.30
0.73
0.76
0.57
0.38
0.24
0.07
0.02

18

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-3.19
-0.35
1.53
2.14
2.17
2.04
1.75
1.31
0.84
0.38

-2.21
0.70
2.56
3.15
3.13
2.85
2.44
1.97
1.37
0.70

-2.96
0.44
2.06
2.57
2.57
2.30
1.90
1.51
0.99
0.48

-3.46
-0.26
1.59
2.01
1.96
1.75
1.44
1.05
0.73
0.38

-4.79
-1.62
0.25
0.66
0.69
0.53
0.35
0.20
-0.06
-0.03

20

0.95
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

-2.71
0.01
1.79
2.43
2.42
2.24
1.89
1.47
1.05
0.52

-1.77
0.96
2.82
3.32
3.22
3.01
2.62
2.08
1.50
0.81

-2.68
0.68
2.21
2.64
2.60
2.40
1.94
1.53
1.07
0.50

-3.22
-0.03
1.63
2.06
1.96
1.72
1.47
1.06
0.77
0.39

-4.36
-1.29
0.36
0.83
0.76
0.57
0.37
0.14
0.01
-0.16

References
Abdulkadiroğlu, A., P. A. Pathak, and A. E. Roth (2009). Strategy-proofness versus efficiency in matching with indifferences: Redesigning the nyc high school
match. American Economic Review 99(5), 1954–78.
Abdulkadiroğlu, A. and T. Sönmez (2003). School choice: A mechanism design
approach. American Economic Review 93(3), pp. 729–747.
Abdulkadiroglu, A. and T. Sönmez (2010). Matching markets: Theory and practice.
Prepared for the Econometric Society World Congress, China (2010).
Adar, E. and B. A. Huberman (2000). Free riding on gnutella. First Monday 5(10).
Alcalde, J. and S. Barberà (1994). Top dominance and the possibility of strategyproof stable solutions to matching problems. Economic Theory 4(3), 417–435.
Aldershof, B. and O. M. Carducci (1999). Stable marriage and genetic algorithms:
A fertile union. Journal of Heuristics 5(1), 29–46.
Anagnostakis, K. G. and M. B. Greenwald (2004). Exchange-based incentive mechanisms for peer-to-peer file sharing. In Distributed Computing Systems, 2004. Proceedings. 24th International Conference on, pp. 524–533. IEEE.
Anderson, D. and G. Fedak (2006). The computational and storage potential of
volunteer computing. In Cluster Computing and the Grid, 2006. CCGRID 06. Sixth
IEEE International Symposium on, Volume 1, pp. 73 –80.
Anderson, D. P. (2004). Boinc: A system for public-resource computing and storage.
In Grid Computing, 2004. Proceedings. Fifth IEEE/ACM International Workshop on,
pp. 4–10. IEEE.
Anderson, D. P., J. Cobb, E. Korpela, M. Lebofsky, and D. Werthimer (2002).
Seti@ home: an experiment in public-resource computing. Communications of
the ACM 45(11), 56–61.

261

262

References

Andreoni, J. and J. Miller (2002). Giving according to garp: An experimental test of
the consistency of preferences for altruism. Econometrica 70(2), pp. 737–753.
Antikainen, M. (2011). Facilitating customer involvement in collaborative online innovation communities. Ph. D. thesis, Tampere University of Technology.
Antoniadis, P. (2004). Comparing economic incentives in peer-to-peer networks.
Computer Networks 46(1), 133–146.
Ariely, D., A. Bracha, and S. Meier (2009). Doing good or doing well? image motivation and monetary incentives in behaving prosocially. American Economic
Review 99(1), 544–555.
Armbrust, M., A. Fox, R. Griffith, A. D. Joseph, R. Katz, A. Konwinski, G. Lee,
D. Patterson, A. Rabkin, I. Stoica, et al. (2010). A view of cloud computing. Communications of the ACM 53(4), 50–58.
Arrow, K. J. (1972). Gifts and exchanges. Philosophy and Public Affairs 1(4), 343–362.
Ashlagi, I. and F. Klijn (2012). Manipulability in matching markets: conflict and
coincidence of interests. Social Choice and Welfare 39(1), 23–33.
Asvanund, A., K. Clay, R. Krishnan, and M. D. Smith (2004). An empirical analysis of network externalities in peer-to-peer music-sharing networks. Information
Systems Research 15(2), 155–174.
Axelrod, R. (1997). The evolution of strategies in the iterated prisoner’s dilemma.
In The dynamics of norms, pp. 1–16.
Axelrod, R. and W. Hamilton (1981).
ence 211(4489), 1390–1396.

The evolution of cooperation.

Sci-

Axtell, R. L. and S. O. Kimbrough (2008). The high cost of stability in two-sided
matching: How much social welfare should be sacrificed in the pursuit of stability? In Proceedings of the 2008 World Congress on Social Simulation (WCSS-08).
Babaoglu, O., M. Marzolla, and M. Tamburini (2012). Design and implementation
of a p2p cloud system. In Proceedings of the 27th Annual ACM Symposium on Applied Computing, pp. 412–417. ACM.
Backstrom, L., P. Boldi, M. Rosa, J. Ugander, and S. Vigna (2012). Four degrees
of separation. In Proceedings of the 3rd Annual ACM Web Science Conference, pp.
33–42. ACM.

References

263

Balinski, M. and T. Sönmez (1999). A tale of two mechanisms: student placement.
Journal of Economic Theory 84(1), 73–94.
Baumgart, I., B. Heep, and S. Krause (2007). Oversim: A flexible overlay network
simulation framework. In IEEE Global Internet Symposium, 2007, pp. 79–84.
Belk, R. (2010). Sharing. Journal of Consumer Research 36(5), pp. 715–734.
Belk, R. (2013). You are what you can access: Sharing and collaborative consumption online. Journal of Business Research.
Bénabou, R. and J. Tirole (2003). Intrinsic and extrinsic motivation. The Review of
Economic Studies 70(3), 489–520.
Bénabou, R. and J. Tirole (2006). Incentives and prosocial behavior. American Economic Review 96(5), pp. 1652–1678.
Berg, J., J. Dickhaut, and K. McCabe (1995). Trust, Reciprocity, and Social History.
Games and Economic Behavior 10(1), 122–142.
Bishop, J. (2007). Increasing participation in online communities: A framework for
human-computer interaction. Computers in Human Behavior 23(4), 1881–1893.
Bolton, G. and A. Ockenfels (2000). ERC: A theory of equity, reciprocity, and competition. American Economic Review 90(1), 166–193.
Botsman, R. and R. Rogers (2010). What’s mine is yours: The rise of collaborative consumption. HarperCollins.
Boudreau, J. (2011). A note on the efficiency and fairness of decentralized matching.
Operations Research Letters 39(4), 231–233.
Boudreau, J. W. and V. Knoblauch (2010). Marriage matching and intercorrelation
of preferences. Journal of Public Economic Theory 12(3), 587–602.
Buyya, R. and M. Murshed (2002). Gridsim: A toolkit for the modeling and simulation of distributed resource management and scheduling for grid computing.
Concurrency and Computation: Practice and Experience (CCPE) 14(13), 1175–1220.
Calheiros, R. N., R. Ranjan, A. Beloglazov, C. A. F. De Rose, and R. Buyya (2011).
Cloudsim: a toolkit for modeling and simulation of cloud computing environments and evaluation of resource provisioning algorithms. Software: Practice and
Experience 41(1), 23–50.

264

References

Cappos, J., I. Beschastnikh, A. Krishnamurthy, and T. Anderson (2009). Seattle: a
platform for educational cloud computing. In ACM SIGCSE Bulletin, Volume 41,
pp. 111–115. ACM.
Casanova, H., A. Legrand, and M. Quinson (2008). Simgrid: A generic framework for large-scale distributed experiments. In Computer Modeling and Simulation, 2008. UKSIM 2008. Tenth International Conference on, pp. 126–131. IEEE.
Caton, S., C. Dukat, T. Grenz, C. Haas, M. Pfadenhauer, and C. Weinhardt (2012).
Foundations of Trust: Contextualising Trust in Social Clouds. In 2nd International
Conference on Social Computing and Its Applications (SCA 2012), pp. 424–429.
Caton, S., C. Haas, K. Chard, K. Bubendorfer, and O. Rana (2014). A social compute
cloud: Allocating and sharing infrastructure resources via social networks. IEEE
Transactions on Services Computing (forthcoming).
Celik, O. B. and V. Knoblauch (2007). Marriage matching with correlated preferences. Economics Working Papers, Paper 200716.
Chan, C., M. Bhandar, L.-B. Oh, and H.-C. Chan (2004). Recognition and participation in a virtual community. In System Sciences, 2004. Proceedings of the 37th
Annual Hawaii International Conference on, pp. 1–10.
Chard, K. and K. Bubendorfer (2008). A distributed economic meta-scheduler for
the grid. In Proceedings of the 8th IEEE International Symposium on Cluster Computing and the Grid (CCGRID), Lyon, France, pp. 542 – 547.
Chard, K., K. Bubendorfer, S. Caton, and O. F. Rana (2012). Social cloud computing:
A vision for socially motivated resource sharing. IEEE Transactions on Services
Computing 5(4), 551–563.
Chard, K., S. Caton, O. Rana, and K. Bubendorfer (2010). Social Cloud: Cloud
Computing in Social Networks. In 2010 IEEE 3rd International Conference on Cloud
Computing (CLOUD), pp. 99–106.
Chard, K., S. Caton, O. Rana, and D. Katz (2012). A social content delivery network
for scientific cooperation: Vision, design, and architecture. In High Performance
Computing, Networking, Storage and Analysis (SCC), 2012 SC Companion:, pp. 1058–
1067.
Charness, G. and M. Rabin (2000). Social preferences: Some simple tests and a new
model. Technical report, University of California at Berkeley.

References

265

Cheng, R. and J. Vassileva (2006). Design and evaluation of an adaptive incentive
mechanism for sustained educational online communities. User Modeling and
User-Adapted Interaction 16(3-4), 321–348.
Cime, W. and K. Marzullo (1999). The computational co-op: Gathering clusters into
a metacomputer. In Parallel and Distributed Processing, 1999. 13th International and
10th Symposium on Parallel and Distributed Processing, pp. 160 –166.
Clary, E. G., M. Snyder, R. D. Ridge, J. Copeland, A. A. Stukas, J. Haugen, and
P. Miene (1998). Understanding and assessing the motivations of volunteers: A
functional approach. Journal of Personality and Social Psychology 74(6), 1516.
Coleman, J. S. (1990). Foundations of Social Theory, Volume 69. Harvard University
Press.
Cropanzano, R. and M. S. Mitchell (2005). Social exchange theory: An interdisciplinary review. Journal of Management 31(6), 874–900.
Cuevas, R., M. Kryczka, A. Cuevas, S. Kaune, C. Guerrero, and R. Rejaie (2010). Is
content publishing in bittorrent altruistic or profit-driven? In Proceedings of the
6th International COnference, pp. 11. ACM.
Curry, R., C. Kiddle, N. Markatchev, R. Simmonds, T. Tan, M. Arlitt, and B. Walker
(2008). Facebook Meets the Virtualized Enterprise. In Proceedings of the 12th International IEEE Enterprise Distributed Object Computing Conference (EDOC ’08),
Washington, DC, USA, pp. 286–292. IEEE.
De Jong, K. A. and W. M. Spears (1991). An analysis of the interacting roles of
population size and crossover in genetic algorithms. In Parallel Problem Solving
from Nature, pp. 38–47. Springer.
De Roure, D., C. Goble, and R. Stevens (2009). The design and realisation of the myexperiment virtual research environment for social sharing of workflows. Future
Generation Computer Systems 25(5), 561–567.
Deci, E. L. and R. M. Ryan (1985). Self-Determination. Wiley Online Library.
Deterding, S., D. Dixon, R. Khaled, and L. Nacke (2011). From game design elements to gamefulness: Defining "gamification". In Proceedings of the 15th International Academic MindTrek Conference: Envisioning Future Media Environments,
MindTrek ’11, New York, NY, USA, pp. 9–15. ACM.

266

References

Deterding, S., M. Sicart, L. Nacke, K. O’Hara, and D. Dixon (2011). Gamification.
using game-design elements in non-gaming contexts. In CHI ’11 Extended Abstracts on Human Factors in Computing Systems, CHI EA ’11, New York, NY, USA,
pp. 2425–2428. ACM.
Diebold, F., H. Aziz, M. Bichler, F. Matthes, and A. Schneider (2014). Course allocation via stable matching. Business & Information Systems Engineering 6(2), 97–110.
Dueck, G. and T. Scheuer (1990). Threshold accepting: A general purpose optimization algorithm appearing superior to simulated annealing. Journal of Computational Physics 90(1), 161 – 175.
Duffy, J. (2006). Chapter 19 agent-based models and human subject experiments. In
L. Tesfatsion and K. Judd (Eds.), Handbook of Computational Economics, Volume 2,
pp. 949 – 1011. Elsevier.
Echenique, F., A. J. Wilson, and L. Yariv (2009). Clearinghouses for two-sided
matching: An experimental study. Social Science Working Paper, 1315.
Echenique, F. and L. Yariv (2011). An experimental study of decentralized matching. Technical report, Working paper.
Ehlers, L. (2008). Truncation strategies in matching markets. Mathematics of Operations Research 33(2), 327–335.
Elnaffar, S., Z. Maamar, and Q. Z. Sheng (2013). When clouds start socializing: The
sky model. International Journal of E-Business Research (IJEBR) 9(2), 1–7.
Elnozahy, E. M., M. Kistler, and R. Rajamony (2003). Energy-efficient server clusters. In Power-Aware Computer Systems, pp. 179–197. Springer.
Emerson, R. M. (1976). Social exchange theory. Annual Review of Sociology 2, 335–
362.
Endress, M. (2012). Trust and the dialectic of the familiar and unfamiliar within the
life-world. Interaction and everyday life: Phenomenological and Ethnomethodological
essays in honor of George Psathas, 115–133.
Erdil, A. and H. Ergin (2006). Two-sided matching with indifferences. Technical
report, Harvard Business School.
Erdil, A. and H. Ergin (2008). What’s the matter with tie-breaking? Improving
efficiency in school choice. American Economic Review 98(3), 669–689.

References

267

Falk, A. and U. Fischbacher (2006). A theory of reciprocity. Games and Economic
Behavior 54(2), 293 – 315.
Feder, T. (1992). A new fixed point approach for stable networks and stable marriages. Journal of Computer and System Sciences 45(2), 233 – 284.
Fehr, E. and U. Fischbacher (2003). The nature of human altruism. Nature 425(6960),
785–791.
Fehr, E. and K. Schmidt (2000). Theories of fairness and reciprocity - evidence
and economic applications. Technical report, Institute for Empirical Research in
Economics - University of Zurich.
Fehr, E. and K. M. Schmidt (2006). The economics of fairness, reciprocity and altruism - experimental evidence and new theories. In S.-C. Kolm and J. M. Ythier
(Eds.), Handbook on the Economics of Giving, Reciprocity and Altruism, Volume 1
of Handbook on the Economics of Giving, Reciprocity and Altruism, Chapter 8, pp.
615–691. Elsevier.
Feldman, M., K. Lai, I. Stoica, and J. Chuang (2004). Robust incentive techniques
for peer-to-peer networks. In Proceedings of the 5th ACM Conference on Electronic
Commerce - EC ’04, pp. 102. ACM.
Festinger, L. (1954). A theory of social comparison processes. Human Relations 7(2),
117–140.
Foa, E. B. and U. G. Foa (2012). Resource theory of social exchange. In Handbook of
Social Resource Theory, pp. 15–32. Springer.
Fogg, B. and D. Eckles (2007). The behavior chain for online participation: How
successful web services structure persuasion. In Y. Kort, W. IJsselsteijn, C. Midden, B. Eggen, and B. Fogg (Eds.), Persuasive Technology, Volume 4744 of Lecture
Notes in Computer Science, pp. 199–209. Springer Berlin Heidelberg.
Foster, I. (2011). Globus online: Accelerating and democratizing science through
cloud-based services. IEEE Internet Computing 15(3), 70–73.
Foster, I. and C. Kesselman (2003). The Grid 2: Blueprint for a new computing infrastructure. Elsevier.
Frey, B. S. and R. Jegen (2001). Motivation crowding theory. Journal of Economic
Surveys 15(5), 589–611.

268

References

Gale, D. and L. Shapley (1962). College admissions and the stability of marriage.
American Mathematical Monthly 69(1), 9–15.
Ganley, D. and C. Lampe (2009). The ties that bind: Social network principles in
online communities. Decision Support Systems 47(3), 266 – 274.
Garland, R. (1991). The mid-point on a rating scale: Is it desirable. Marketing
Bulletin 2(1), 66–70.
Gelain, M., M. S. Pini, F. Rossi, K. B. Venable, and T. Walsh (2013). Local search
approaches in stable matching problems. Algorithms 6(4), 591–617.
Gent, I. P. and P. Prosser (2002). An empirical study of the stable marriage problem
with ties and incomplete lists. In ECAI - 15th European Conference on Artificial
Intelligence, pp. 141–145.
Golbeck, J. A. (2005). Computing and Applying Trust in Web-based Social Networks. Ph.
D. thesis, University of Maryland.
Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley Longman Publishing Co., Inc.
Golle, P., K. Leyton-Brown, I. Mironov, and M. Lillibridge (2001). Incentives for
sharing in peer-to-peer networks. In Electronic Commerce, pp. 75–87. Springer.
Gosling, S. D., P. J. Rentfrow, and W. B. Swann Jr (2003). A very brief measure of
the big-five personality domains. Journal of Research in Personality 37(6), 504–528.
Gracia-Tinedo, R., M. Sanchez-Artigas, and P. Garcia-Lopez (2012). F2box: Cloudifying f2f storage systems with high availability correlation. In Cloud Computing
(CLOUD), 2012 IEEE 5th International Conference on, pp. 123–130.
Granovetter, M. (1985). Economic action and social structure: The problem of embeddedness. American Journal of Sociology 91(3), 481–510.
Granovetter, M. (2005). The Impact of Social Structure on Economic Outcomes.
Journal of Economic Perspectives 19(1), 33–50.
Granovetter, M. S. (1973). The strength of weak ties. The American Journal of Sociology 78(6), 1360–1380.
Guo, Z., R. Singh, and M. Pierce (2009). Building the polargrid portal using web 2.0
and opensocial. In Proceedings of the 5th Grid Computing Environments Workshop
(GCE ’09), New York, NY, USA, pp. 1–8. ACM.

References

269

Gusfield, D. (1987). Three fast algorithms for four problems in stable marriage.
SIAM Journal on Computing 16(1), 111–128.
Gusfield, D. and R. W. Irving (1989). The Stable Marriage Problem: Structure and
Algorithms. Cambridge, MA: MIT Press.
Haas, C., S. Caton, K. Chard, and C. Weinhardt (2013). Co-Operative Infrastructures: An Economic Model for Providing Infrastructures for Social Cloud Computing. In Proceedings of the Forty-Sixth Annual Hawaii International Conference on
System Sciences (HICSS). (Grand Wailea, Maui, USA), pp. 729–738.
Haas, C., S. Caton, D. Trumpp, and C. Weinhardt (2012). A Simulator for Social
Exchanges and Collaborations - Architecture and Case Study. In Proceedings of
the 8th IEEE International Conference on eScience (eScience 2012).
Haas, C., S. Caton, and C. Weinhardt (2011). Engineering incentives in social
clouds. In Proceedings of the 11th IEEE/ACM International Symposium on Cluster,
Cloud and Grid Computing (CCGrid 2011), pp. 572–575.
Haas, C., S. O. Kimbrough, S. Caton, and C. Weinhardt (2013). Preference-based
resource allocation: Using heuristics to solve two-sided matching problems with
indifferences. In J. Altmann, K. Vanmechelen, and O. Rana (Eds.), GECON 2013,
LNCS 8193, pp. 149–160. Springer International Publishing Switzerland.
Haas, C., S. O. Kimbrough, and C. van Dinther (2013). Strategic learning by eservice suppliers in service value networks. Journal of Service Research 16(3), 259
– 276.
Halldórsson, M., K. Iwama, S. Miyazaki, and H. Yanagisawa (2007). Improved
approximation results for the stable marriage problem. ACM Transactions on Algorithms (TALG) 3(3), 30. Article 30.
Halldórsson, M. M., R. W. Irving, K. Iwama, D. F. Manlove, S. Miyazaki, Y. Morita,
and S. Scott (2003). Approximability results for stable marriage problems with
ties. Theoretical Computer Science 306(1-3), 431 – 447.
Harper, F. M., S. X. Li, Y. Chen, and J. A. Konstan (2007). Social comparisons to
motivate contributions to an online community. In Persuasive Technology, pp.
148–159. Springer.

270

References

Hars, A. and S. Ou (2001). Working for free? Motivations of participating in open
source projects. In System Sciences, 2001. Proceedings of the 34th Annual Hawaii
International Conference on.
Häsel, M. (2011). Opensocial: an enabler for social applications on the web. Communications of the ACM 54(1), 139–144.
Hearne, R. (2006). SERP click through rate of google search results. Red Cardinal.
August 12, 2006. http://www.redcardinal.ie/google/12-08-2006/clickthroughanalysis-of-aol-datatgz/ – last accessed April 2014.
Henrich, J., R. Boyd, S. Bowles, C. Camerer, E. Fehr, H. Gintis, and R. McElreath
(2001). In search of homo economicus: Behavioral experiments in 15 small-scale
societies. American Economic Review 91(2), pp. 73–78.
Holland, J. H. (1992). Adaptation in natural and artificial systems. MIT Press.
Inamura, T., T. Shibata, H. Sena, T. Hashimoto, N. Kawai, T. Miyashita, Y. Sakurai,
M. Shimizu, M. Otake, K. Hosoda, et al. (2010). Simulator platform that enables
social interaction simulation - sigverse: Sociointelligenesis simulator. In System
Integration (SII), 2010 IEEE/SICE International Symposium on, pp. 212–217. IEEE.
Iriberri, A. and G. Leroy (2009). A life-cycle perspective on online community success. ACM Computing Surveys (CSUR) 41(2), 11:1–11:29.
Irving, R. and P. Leather (1986). The complexity of counting stable marriages. SIAM
Journal on Computing 15(3), 655–667.
Irving, R. W. (1994). Stable marriage and indifference. Discrete Applied Mathematics 48(3), 261 – 272.
Irving, R. W., P. Leather, and D. Gusfield (1987). An efficient algorithm for the
optimal stable marriage. Journal of the ACM 34(3), 532–543.
Irving, R. W. and D. F. Manlove (2008). Approximation algorithms for hard variants
of the stable marriage and hospitals/residents problems. Journal of Combinatorial
Optimization 16(3), 279–292.
Irwin, D., J. Chase, L. Grit, and A. Yumerefendi (2005). Self-recharging virtual
currency. In Proceedings of the 2005 ACM SIGCOMM Workshop on Economics of
Peer-to-peer Systems, P2PECON ’05, New York, NY, USA, pp. 93–98. ACM.

References

271

Iwama, K., S. Miyazaki, Y. Morita, and D. Manlove (1999). Stable marriage with
incomplete lists and ties. In J. Wiedermann, P. Emde Boas, and M. Nielsen (Eds.),
Automata, Languages and Programming, Volume 1644 of Lecture Notes in Computer
Science, pp. 443–452. Springer Berlin Heidelberg.
Iwama, K., S. Miyazaki, and H. Yanagisawa (2010). Approximation algorithms
for the sex-equal stable marriage problem. ACM Transactions on Algorithms
(TALG) 7(1), 2.
Iwama, K., S. Miyazaki, and H. Yanagisawa (2014). A 25/17-approximation algorithm for the stable marriage problem with one-sided ties. Algorithmica 68(3),
758–775.
Jackson, M. O. (2003). Mechanism theory. In U. Derigs (Ed.), Encyclopedia of Life
Support Systems. EOLSS Publishers: Oxford UK.
Javadi, B., D. Kondo, J. Vincent, and D. Anderson (2011). Discovering statistical models of availability in large distributed systems: An empirical study of
seti@home. Parallel and Distributed Systems, IEEE Transactions on PP(99), 1.
Javadi, B., D. Kondo, J.-M. Vincent, and D. P. Anderson (2009). Mining for statistical
models of availability in large-scale distributed systems: An empirical study of
seti@ home. In Modeling, Analysis & Simulation of Computer and Telecommunication
Systems, 2009. MASCOTS’09. IEEE International Symposium on, pp. 1–10. IEEE.
John, K., K. Bubendorfer, and K. Chard (2011). A social cloud for public eresearch.
In E-Science (e-Science), 2011 IEEE 7th International Conference on, pp. 363 –370.
John, O. P., E. M. Donahue, and R. L. Kentle (1991). The big five inventory - versions
4a and 54. Berkeley: University of California, Berkeley, Institute of Personality and
Social Research.
Jones, Q. and S. Rafaeli (1999). User population and user contributions to virtual
publics: a systems model. In Proceedings of the International ACM SIGGROUP
Conference on Supporting Group Work, GROUP ’99, New York, NY, USA, pp. 239–
248. ACM.
Jøsang, A., R. Ismail, and C. Boyd (2007). A survey of trust and reputation systems
for online service provision. Decision Support Systems 43(2), 618–644.

272

References

Kagel, J. H. and A. E. Roth (2000). The dynamics of reorganization in matching markets: A laboratory experiment motivated by a natural experiment. The Quarterly
Journal of Economics 115(1), 201–235.
Kamvar, S. D., M. T. Schlosser, and H. Garcia-molina (2003). The Eigen-Trust Algorithm for Reputation Management in P2P Networks. In Proceedings of the 12th
International World Wide Web Conference, Budapest, Hungary, pp. 640–651.
Karau, S. J. and K. D. Williams (1993). Social loafing: A meta-analytic review and
theoretical integration. Journal of Personality and Social Psychology 65(4), 681.
Kesten, O. (2012). On two kinds of manipulation for school choice problems. Economic Theory 51(3), 677–693.
Khan, A. M., L. Navarro, L. Sharifi, and L. Veiga (2013). Clouds of small things: Provisioning infrastructure-as-a-service from within community networks. In Wireless and Mobile Computing, Networking and Communications (WiMob), 2013 IEEE
9th International Conference on, pp. 16–21. IEEE.
Kimbrough, S. and A. Kuo (2010). On heuristics for two-sided matching: revisiting
the stable marriage problem as a multiobjective problem. In Proceedings of the
12th Annual Conference on Genetic and Evolutionary Computation, pp. 1283–1290.
ACM.
Kimbrough, S. O. and F. H. Murphy (2009). Learning to Collude Tacitly on Production Levels by Oligopolistic Agents. Computational Economics 33(1), 47–78.
Király, Z. (2011). Better and simpler approximation algorithms for the stable marriage problem. Algorithmica 60(1), 3–20.
Klaus, B. and F. Klijn (2006a). Median stable matching for college admissions. International Journal of Game Theory 34(1), 1–11.
Klaus, B. and F. Klijn (2006b). Procedurally fair and stable matching. Economic
Theory 27, 431–447.
Klimeck, G., M. McLennan, S. P. Brophy, G. B. Adams, and M. S. Lundstrom (2008).
nanoHUB.org: Advancing Education and Research in Nanotechnology. Computing in Science and Engineering 10, 17–23.
Knuth, D. E. (1976). Marriages Stables.
l’Université de Montreal.

Montreal, Canada: Les Presses de

References

273

Knuth, D. E. (1997). Stable Marriage and Its Relation to Other Combinatorial Problems:
An Introduction to the Mathematical Analysis of Algorithms, Volume 10 of CRM Proceedings & Lecture Notes, Centre de Recherches Mathématiques Université de Montréal.
Providence, RI: American Mathematical Society. Originally published as Knuth
(1976).
Koh, J., Y. Kim, B. Butler, and G. Bock (2007). Encouraging participation in virtual
communities. Communications of the ACM 50(2), 68–73.
Kojima, F. and P. A. Pathak (2009). Incentives and stability in large two-sided
matching markets. American Economic Review 99(3), pp. 608–627.
Kollock, P. (1999). The economies of online cooperation: Gifts and public goods
in cyberspace. In P. Kollock and M. Smith (Eds.), Communities in Cyberspace, pp.
219–240. Routledge.
Kourtellis, N. (2012). On the Design of Socially-Aware Distributed Systems. Ph. D.
thesis, University of South Florida.
Kourtellis, N., J. Finnis, P. Anderson, J. Blackburn, C. Borcea, and A. Iamnitchi
(2010). Prometheus: User-controlled p2p social data management for sociallyaware applications. In I. Gupta and C. Mascolo (Eds.), Middleware 2010, Volume
6452 of Lecture Notes in Computer Science, pp. 212–231. Springer Berlin Heidelberg.
Krishnan, R., M. Smith, and R. Telang (2006). The economics of peer-to-peer networks. Journal of Information Technology Theory and Application (JITTA) 5, 31–44.
Lakhani, K. R. and E. von Hippel (2003). How open source software works: “free”
user-to-user assistance. Research Policy 32(6), 923 – 943.
Lakhani, K. R. and R. G. Wolf (2005). Why hackers do what they do: Understanding motivation and effort in Free/Open source software projects. In J. Feller,
B. Fitzgerald, S. A. Hissam, and K. R. Lakhani (Eds.), Perspectives on Free and
Open Source Software, pp. 3–21. Cambridge, Massachusetts: The MIT Press.
Lampe, C., R. Wash, A. Velasquez, and E. Ozkaya (2010). Motivations to participate
in online communities. In Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems, CHI ’10, New York, NY, USA, pp. 1927–1936. ACM.
Law, A. M. (2007). Simulation modeling and analysis (4. ed. ed.). McGraw-Hill series
in industrial engineering and management science. McGraw-Hill.

274

References

Lazar, J. and J. Preece (2002). Social Considerations in Online Communities: Usability, Sociability, and Success Factors. In H. Van Oostendorp (Ed.), Cognition in
a Digital World. Mahwah: NJ: Lawrence Erlbaum Associates Inc. Publishers.
Lee, S. (2011). Incentive compatibility of large centralized matching markets. Job
market paper.
Levine, M. (2009). Share my ride. New York Times Magazine. March 9, 2009.
Liao, R. R.-F., R. H. Wouhaybi, and A. T. Campbell (2002). Incentive engineering
in wireless LAN based access networks. In 10th IEEE International Conference on
Network Protocols, 2002. Proceedings., pp. 58–67. IEEE.
Manlove, D. F., R. W. Irving, K. Iwama, S. Miyazaki, and Y. Morita (2002). Hard
variants of stable marriage. Theoretical Computer Science 276(1), 261 – 279.
Mao, Y., J. Vassileva, and W. Grassmann (2007). A system dynamics approach to
study virtual communities. In System Sciences, 2007. HICSS 2007. 40th Annual
Hawaii International Conference on.
Markus, M. L. (1987). Toward a "critical mass" theory of interactive media: Universal access, interdependence and diffusion. Communication Research 14(5), 491–511.
Mas-Colell, A., M. D. Whinston, J. R. Green, et al. (1995). Microeconomic Theory,
Volume 1. Oxford university press, New York.
McDermid, E. (2009). A 3/2-approximation algorithm for general stable marriage.
In Automata, Languages and Programming, pp. 689–700. Springer.
Mell, P. and T. Grance (2011). The nist definition of cloud computing. Technical Report 800-145, National Institute of Standards and Technology (NIST), Gaithersburg, MD.
Michalk, W. and C. Haas (2011). Incentives in service level agreement establishment the case of economic and social aspects. In Requirements Engineering for
Social Computing (RESC), 2011 First International Workshop on, pp. 30–33.
Mohaisen, A., H. Tran, A. Chandra, and Y. Kim (2011). SocialCloud: Using Social Networks for Building Distributed Computing Services. Technical report,
Cornell University.

References

275

Montresor, A. and M. Jelasity (2009). Peersim: A scalable p2p simulator. In Peer-toPeer Computing, 2009. P2P’09. IEEE Ninth International Conference on, pp. 99–100.
IEEE.
Mowbray, M., F. Brasileiro, N. Andrade, J. Santana, and W. Cirne (2006). A
Reciprocation-Based Economy for Multiple Services in Peer-to-Peer Grids. In
Sixth IEEE International Conference on Peer-to-Peer Computing (P2P’06), pp. 193–
202. IEEE.
Muck, P. M., B. Hell, and S. D. Gosling (2007). Construct validation of a short
five-factor model instrument: A self-peer study on the german adaptation of the
ten-item personality inventory (tipi-g). European Journal of Psychological Assessment 23(3), 166.
Nakamura, M., K. Onaga, S. Kyan, and M. Silva (1995). Genetic algorithm for sexfair stable marriage problem. In Circuits and Systems, 1995. ISCAS ’95, 1995 IEEE
International Symposium on, Volume 1, pp. 509 –512 vol.1.
Neumann, D. G. (2004). Market Engineering: A structured design process for electronic
markets. Univ.-Verlag Karlsruhe.
Nisan, N. and A. Ronen (1999). Algorithmic mechanism design. In Proceedings of the
Thirty-First Annual ACM Symposium on Theory of Computing, pp. 129–140. ACM.
Nobel Prize Committee (2012). Alvin E. Roth and Lloyd S. Shapley: Stable allocations and the practice of market design. Technical report, Royal Swedish
Academy of Sciences.
Nov, O. (2007). What motivates wikipedians? Communications of the ACM 50(11),
60–64.
Nov, O., D. Anderson, and O. Arazy (2010). Volunteer computing: a model of
the factors determining contribution to community-based scientific research. In
Proceedings of the 19th International Conference on World Wide Web, WWW ’10, New
York, NY, USA, pp. 741–750. ACM.
Nov, O., M. Naaman, and C. Ye (2010). Analysis of participation in an online photosharing community: A multidimensional perspective. Journal of the American
Society for Information Science and Technology 61, 555–566.
Odlyzko, A. (2003). The case against micropayments. In R. Wright (Ed.), Financial
Cryptography, pp. 77–83. Springer.

276

References

Pais, J. (2008). Random matching in the college admissions problem. Economic
Theory 35(1), 99–116.
Paluch, K. (2012). Faster and simpler approximation of stable matchings. In Approximation and Online Algorithms, pp. 176–187. Springer.
Parkes, D. C. (2001). Iterative combinatorial auctions: Achieving economic and computational efficiency. Ph. D. thesis, University of Pennsylvania.
Pavlou, P. A. and D. Gefen (2004). Building effective online marketplaces with
institution-based trust. Information Systems Research 15(1), 37–59.
Petri, I., O. F. Rana, Y. Regzui, and G. C. Silaghi (2012). Risk assessment in service
provider communities. In Economics of Grids, Clouds, Systems, and Services, pp.
135–147. Springer.
Petri, I., O. F. Rana, Y. Rezgui, and G. C. Silaghi (2012). Trust modelling and analysis
in peer–to–peer clouds. International Journal of Cloud Computing 1(2), 221–239.
Pini, M. S., F. Rossi, K. B. Venable, and T. Walsh (2011a). Manipulation complexity
and gender neutrality in stable marriage procedures. Autonomous Agents and
Multi-Agent Systems 22(1), 183–199.
Pini, M. S., F. Rossi, K. B. Venable, and T. Walsh (2011b). Weights in stable marriage
problems increase manipulation opportunities. In Proceedings of the 13th Conference on Theoretical Aspects of Rationality and Knowledge, TARK XIII, New York, NY,
USA, pp. 200–204. ACM.
Porter, J. (2010). Designing for the social web. Peachpit Press.
Porter, P. K. and G. W. Scully (1987). Economic efficiency in cooperatives. Journal of
Law and Economics 30(2), pp. 489–512.
Preece, J. and D. Maloney-Krichmar (2003). Online communities: focusing on sociability and usability. Lawrence Erlbaum Publishers, Mahwak: NJ.
Priedhorsky, R., J. Chen, S. T. K. Lam, K. Panciera, L. Terveen, and J. Riedl (2007).
Creating, destroying, and restoring value in wikipedia. In Proceedings of the 2007
International ACM Conference on Supporting Group Work, GROUP ’07, New York,
NY, USA, pp. 259–268. ACM.

References

277

Rafaeli, S. and Y. Ariel (2008). Online motivational factors: Incentives for participation and contribution in wikipedia. In A. Barak (Ed.), Psychological Aspects of
Cyberspace. New York, NY: Cambridge University Press.
Ranganathan, K., M. Ripeanu, A. Sarin, and I. Foster (2004). Incentive mechanisms
for large collaborative resource sharing. In IEEE International Symposium on Cluster Computing and the Grid, 2004. CCGrid 2004, pp. 1–8. IEEE.
Ranjan, R., L. Zhao, X. Wu, A. Liu, A. Quiroz, and M. Parashar (2010). Peer-to-peer
cloud provisioning: Service discovery and load-balancing. In Cloud Computing,
pp. 195–217. Springer.
Ren, Y., R. Kraut, and S. Kiesler (2007). Applying common identity and bond theory
to design of online communities. Organization Studies 28(3), 377–408.
Ren, Y. and R. E. Kraut (2011). A simulation for designing online community: Member motivation, contribution, and discussion moderation. Unpublished.
Rheingold, H. (1993). The virtual community. Homesteading on the electronic frontier.
New York,: Addison-Wesley.
Rifkin, J. (2000). The age of access : the new culture of hypercapitalism, where all of life is
a paid-for experience. New York, NY: Jeremy P. Tarcher/Putnam.
Romero-Medina, A. (2001). Sex-equal’stable matchings. Theory and Decision 50(3),
197–212.
Roth, A. (1982). The economics of matching: Stability and incentives. Mathematics
of Operations Research 7(4), 617–628.
Roth, A. (2008). Deferred acceptance algorithms: History, theory, practice, and
open questions. International Journal of Game Theory 36(3), 537–569.
Roth, A. and M. Sotomayor (1992). Two-sided matching: A study in game-theoretic
modeling and analysis, Volume 18. Cambridge University Press.
Roth, A. E. (1985). The college admissions problem is not equivalent to the marriage
problem. Journal of Economic Theory 36(2), 277–288.
Roth, A. E. and U. G. Rothblum (1999). Truncation strategies in matching marketsin search of advice for participants. Econometrica 67(1), pp. 21–43.

278

References

Roth, A. E., U. G. Rothblum, and J. H. V. Vate (1993). Stable matchings, optimal
assignments, and linear programming. Mathematics of Operations Research 18(4),
pp. 803–828.
Rothblum, U. G. (1992). Characterization of stable matchings as extreme points of
a polytope. Mathematical Programming 54(1-3), 57–67.
Roure, D. D., C. Goble, and R. Stevens (2009). The design and realisation of the myexperiment virtual research environment for social sharing of workflows. Future
Generation Computer Systems 25(5), 561–567.
Ryan, R. M. and E. L. Deci (2000). Intrinsic and extrinsic motivations: Classic definitions and new directions. Contemporary Educational Psychology 25(1), 54 – 67.
Sabater, J. and C. Sierra (2002). Social ReGreT, A reputation model based on social
relations. ACM SIGecom Exchanges 3(1), 44–56.
Schnell, R., P. B. Hill, and E. Esser (2011). Methoden der empirischen Sozialforschung.
Oldenbourg Verlag.
Scott, S. (2005). A study of stable marriage problems with ties. Ph. D. thesis, University
of Glasgow.
Seuken, S., D. Charles, M. Chickering, and S. Puri (2010). Market design & analysis
for a p2p backup system. In Proceedings of the 11th ACM Conference on Electronic
Commerce, pp. 97–108. ACM.
Seuken, S., K. Jain, D. Tan, and M. Czerwinski (2010). Hidden markets: Ui design
for a p2p backup application. In Proceedings of the 28th International Conference on
Human Factors in Computing Systems, pp. 315–324. ACM.
Sexton, R. J. (1986). The formation of cooperatives: A game-theoretic approach with
implications for cooperative finance, decision making, and stability. American
Journal of Agricultural Economics 68(2), 214–225.
Shampanier, K., N. Mazar, and D. Ariely (2007). Zero as a special price: The true
value of free products. Marketing Science 26(6), 742–757.
Skinner, B. F. (1953). Science and Human Behavior. New York, NY: Macmillan.
Smith, V. L. (1982). Microeconomic systems as an experimental science. American
Economic Review 72(5), 923–955.

References

279

Sönmez, T. (1999). Can pre-arranged matches be avoided in two-sided matching
markets? Journal of Economic Theory 86(1), 148–156.
Spear, R. (2000). The co-operative advantage. Annals of Public and Cooperative Economics 71(4), 507–523.
Stephen, A. T. and O. Toubia (2010). Deriving Value from Social Commerce Networks. Journal of Marketing Research 47(2), 215–228.
Streitberger, W. and T. Eymann (2009). A simulation of an economic, self-organising
resource allocation approach for application layer networks. Computer Networks 53(10), 1760–1770.
Suls, J., R. Martin, and L. Wheeler (2002). Social comparison: Why, with whom,
and with what effect? Current Directions in Psychological Science 11(5), 159–163.
Swamynathan, G., C. Wilson, B. Boe, K. Almeroth, and B. Y. Zhao (2008). Do social
networks improve e-commerce?: a study on social marketplaces. In Proceedings
of the First Workshop on Online Social Networks, WOSN ’08, New York, NY, USA,
pp. 1–6.
Tapscott, D. and A. D. Williams (2007). Wikinomics: How Mass Collaboration Changes
Everything. New York: Penguin.
Tedjamulia, S., D. Dean, D. Olsen, and C. Albrecht (2005). Motivating content contributions to online communities: Toward a more comprehensive theory. In System Sciences, 2005. HICSS ’05. Proceedings of the 38th Annual Hawaii International
Conference on.
Thal, R. (2013). Representing Agreements in Social Clouds. Master’s thesis, Karlsruhe Institute of Technology.
Thaufeeg, A., K. Bubendorfer, and K. Chard (2011). Collaborative eResearch in a
social cloud. In E-Science (e-Science), 2011 IEEE 7th International Conference on, pp.
224 –231.
Thorndike, E. L. (1927). The Law of Effect. American Journal of Psychology 39, 212–
222.
Tran, D. N., F. Chiang, and J. Li (2008). Friendstore: Cooperative online backup using trusted nodes. In Proceedings of the 1st International Workshop on Social Network
Systems (SocialNet’08), Glasgow, Scotland.

280

References

van Dijck, J. (2009). Users like you? Theorizing agency in user-generated content.
Media, Culture & Society 31(1), 41–58.
Van Slyke, C., V. Ilie, H. Lou, and T. Stafford (2007). Perceived critical mass and
the adoption of a communication technology. European Journal of Information Systems 16(3), 270–283.
Vande Vate, J. H. (1989). Linear programming brings marital bliss. Operations Research Letters 8(3), 147 – 153.
Vargo, S. L. and R. F. Lusch (2004). Evolving to a new dominant logic for marketing.
Journal of Marketing 68, 1–17.
Vassileva, J. (2012). Motivating participation in social computing applications: a
user modeling perspective. User Modeling and User-Adapted Interaction 22(1-2),
177–201.
Venkatesh, V., M. G. Morris, G. B. Davis, and F. D. Davis (2003). User acceptance of
information technology: Toward a unified view. MIS Quarterly 27(3), 425–478.
Vien, N. A. and T. C. Chung (2006). Multiobjective fitness functions for stable marriage problem using genetic algrithm. In SICE-ICASE, 2006. International Joint
Conference, pp. 5500 –5503.
Walter, F. E., S. Battiston, and F. Schweitzer (2008). A model of a trust-based recommendation system on a social network. Autonomous Agents and Multi-Agent
Systems 16(1), 57–74.
Wang, H., J. E. Chung, N. Park, M. L. McLaughlin, and J. Fulk (2012). Understanding online community participation: A technology acceptance perspective.
Communication Research 39(6), 781–801.
Watts, D. J. and S. H. Strogatz (1998). Collective dynamics of ‘small-world’ networks. Nature 393(6684), 440–442.
Weijters, B., E. Cabooter, and N. Schillewaert (2010). The effect of rating scale format on response styles: The number of response categories and response category labels. International Journal of Research in Marketing 27(3), 236–247.
Weinhardt, C., C. Holtmann, and D. Neumann (2003).
Wirtschaftsinformatik 45(6), 635–640.

Market-Engineering.

References

281

Westland, J. C. (2010). Critical mass and willingness to pay for social networks.
Electronic Commerce Research and Applications 9(1), 6 – 19. Special Issue: Social
Networks and Web 2.0.
Wigfield, A. and J. S. Eccles (2000). Expectancy-value theory of achievement motivation. Contemporary Educational Psychology 25(1), 68 – 81.
Williams, R. L. and J. Cothrel (2000). Four smart ways to run online communities.
Sloan Management Review 41(4), 81–92.
Xiang, R., J. Neville, and M. Rogati (2010). Modeling relationship strength in online
social networks. In Proceedings of the 19th International Conference on World Wide
Web, WWW ’10, New York, NY, USA, pp. 981–990. ACM.
Xiaohui, W., D. Zhaohui, Y. Shutao, H. Chang, and L. Huizhen (2006). CSF4: A
WSRF Compliant Meta-Scheduler. In In Proceedings of World Congress in Computer
Science, Computer Engineering, and Applied Computing, pp. 61–67.
Xiong, L. and L. Liu (2004). Peertrust: Supporting reputation-based trust for peerto-peer electronic communities. IEEE Transactions On Knowledge and Data Engineering 16(7), 843–857.
Xu, S. and M. Yung (2010). Socialclouds: Concept, security architecture and some
mechanisms. In L. Chen and M. Yung (Eds.), Trusted Systems, Volume 6163 of
LNCS, pp. 104–128. Springer.
Zhang, K., N. Antonopoulos, and Z. Mahmood (2009). A review of incentive mechanism in peer-to-peer systems. In Advances in P2P Systems, 2009. AP2PS ’09. First
International Conference on, pp. 45–50.
Zhou, R. and K. Hwang (2007). Powertrust: A robust and scalable reputation system for trusted peer-to-peer computing. IEEE Transactions on Parallel and Distributed Systems 18(4), 460–473.
Zhuang, Y., A. Rafetseder, and J. Cappos (2013). Experience with Seattle: A Community Platform for Research and Education. In Research and Educational Experiment Workshop (GREE), 2013 Second GENI, pp. 37–44. IEEE.

An unprecedented variety of resources are shared within (social) networks of users.
In this context, the Social Cloud framework leverages existing relationships between
members of a social network for the exchange of resources. This thesis focuses on
the design of coordination mechanisms to address two challenges for the creation of
a sustainable Social Cloud platform: user participation incentives and resource allocation mechanisms. In the first part, based on the survey-based identification of relevant
participation incentives, two case studies show the usefulness of applying simulations
in the engineering of contribution schemes. The second part of the thesis advocates
the use of two-sided matching for resource allocation in Social Clouds. Heuristics for
two-sided matching are proposed and evaluated as a means to calculate high quality
solutions and provide flexibility with respect to diverse preferences and goals.

ISBN 978-3-7315-0237-1

ISSN 1862-8893
ISBN 978-3-7315-0237-1

9 783731 502371