Lecture Notes in Artificial Intelligence 13023
Subseries of Lecture Notes in Computer Science
Series Editors
Randy Goebel
University of Alberta, Edmonton, Canada Yuzuru Tanaka
Hokkaido University, Sapporo, Japan Wolfgang Wahlster
DFKI and Saarland University, Saarbrücken, Germany
Founding Editor
Jörg Siekmann
DFKI and Saarland University, Saarbrücken, Germany
More information about this subseries athttp://www.springer.com/series/1244
Dimitris Fotakis • David Ríos Insua (Eds.)
Algorithmic
Decision Theory
7th International Conference, ADT 2021 Toulouse, France, November 3–5, 2021 Proceedings
123
Editors Dimitris Fotakis
National Technical University of Athens Athens, Greece
David Ríos Insua
Consejo Superior de Investigaciones Cientificas
Madrid, Madrid, Spain
ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Artificial Intelligence
ISBN 978-3-030-87755-2 ISBN 978-3-030-87756-9 (eBook) https://doi.org/10.1007/978-3-030-87756-9
LNCS Sublibrary: SL7– Artificial Intelligence
© Springer Nature Switzerland AG 2021
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The 7th International Conference on Algorithmic Decision Theory (ADT 2021), held in November 2021, at the University of Toulouse 1 Capitole, France, has continued in the tradition established by previous ADT conferences in providing a unique opportunity for scientific exchange among researchers and practitioners coming from diverse areas of computer science, economics, and operations research. Their joint aim is to improve the theory and practice of modern algorithmic decision support. Previous ADT con- ferences were held in Venice, Italy (2009); Piscataway, NJ, USA (2011); Brussels, Belgium (2013); Lexington, KY, USA (2015), Luxembourg (2017) and Durham, NC, USA (2019).
ADT 2021 received 58 submissions, which were all rigorously peer-reviewed by at least three Program Committee (PC) members, in a double-blind fashion. The papers were evaluated on the basis of originality, significance, and exposition. The PC eventually decided to accept 27 papers to be presented at the conference and to be included in the proceedings. The acceptance rate was 46.5%.
The program also included three invited talks by distinguished researchers in algorithmic decision theory, namely Battista Biggio (University of Cagliari, Italy), Edith Elkind (University of Oxford, UK), and Christophe Labreuche (Thales Research and Technology and SINCLAIR AI Lab, France). In addition, ADT 2021 featured a PhD student day, co-chaired by Georgios Amanatidis (University of Essex, UK), Roi Naveiro (Institute of Mathematical Sciences, Spain), and Arianna Novaro (University of Amsterdam, Netherlands).
The works accepted for publication in this volume cover most of the major aspects of algorithmic decision theory, such as preference modeling and elicitation, compu- tational social choice, preference aggregation, voting, fair division and resource allo- cation, coalition formation, stable matchings, and participatory budgeting.
We thank the authors for their interest in submitting and presenting their high quality recent work to ADT 2021, as well as the PC members and the external reviewers for their great work in evaluating the submissions. We also want to thank the Artificial Intelligence Journal, the EURO Working Group on Preference Handling, the University of Toulouse 1 Capitole, the European Office of Aerospace Research and Development (EOARD), and the AXA Research Fund (through the AXA-ICMAT Chair in Adversarial Risk Analysis) for their generous financial support. We are grateful to the University of Toulouse 1 Capitole for hosting ADT 2021. Special thanks also go to the members of the local organizing committee, Umberto Grandi (chair), Sylvie Doutre, Laurent Perrussel, and Pascale Zaraté, for their excellent organization and local arrangements work, and to Rachael Colley for her help with the conference website. Finally, we want to thank Alexis Tsoukiàs, for his invaluable advice and
support, Anna Kramer at Springer for helping with the proceedings, and the EasyChair conference management system.
August 2021 Dimitris Fotakis
David Ríos Insua
vi Preface
Organization
Program Committee
Haris Aziz The University of New South Wales and Data61, CSIRO, Australia
Katarina Cechlarova Pavol JozefŠafárik University, Slovakia
Lea Deleris BNP Paribas, France
Luis Dias University of Coimbra, Portugal
Love Ekenberg International Institute of Applied Systems Analysis, Austria
Ulle Endriss University of Amsterdam, The Netherlands
Piotr Faliszewski AGH University of Science and Technology, Poland
Angelo Fanelli CNRS, CREM, France
Aris Filos-Ratsikas University of Liverpool, UK
Dimitris Fotakis (Co-chair) National Technical University of Athens, Greece Laurent Gourves CNRS, LAMSADE, Université Paris-Dauphine, France Tatiana V. Guy Institute of Information Theory and Automation,
Czech Academy of Sciences, Czech Republic Carlos Henggeler Antunes University of Coimbra, Portugal
Maria Kyropoulou University of Essex, UK
Jérôme Lang CNRS, LAMSADE, Université Paris-Dauphine, France David Manlove University of Glasgow, UK
Evangelos Markakis Athens University of Economics and Business, Greece Reshef Meir Technion-Israel Institute of Technology, Israel Fanny Pascual LIP6, Université Pierre et Marie Curie - Paris 6, France Patrice Perny LIP6, Université Pierre et Marie Curie - Paris 6, France Hans Peters Maastricht University, The Netherlands
Marc Pirlot Université de Mons, Belgium Maria Polukarov King’s College London, UK David Rios Insua (Co-chair) Universidad Rey Juan Carlos, Spain Fred Roberts Rutgers University, USA
Francesca Rossi IBM Research, USA Antonio Salmeron University of Almería, Spain
Ahti Salo Aalto University School of Science and Technology, Finland
Maria Serna Universitat Politècnica de Catalunya, Spain
Alexis Tsoukias CNRS, LAMSADE, Université Paris-Dauphine, France Carmine Ventre King’s College London, UK
Toby Walsh The University of New South Wales, Australia
Organizing Committee
Rachael Colley University of Toulouse 1 Capitole, France Sylvie Doutre University of Toulouse 1 Capitole, France Umberto Grandi (Chair) University of Toulouse 1 Capitole, France Laurent Perrussel University of Toulouse 1 Capitole, France Pascale Zaraté University of Toulouse 1 Capitole, France
Additional Reviewers
Amanatidis, Georgios Archbold, Thomas Auletta, Vincenzo Bachrach, Yoram Bielous, Gili Boixel, Arthur Bouveret, Sylvain Bredereck, Robert Cailloux, Olivier Eirinakis, Pavlos Fairstein, Roy Ferraioli, Diodato Greco, Gianluigi Hamm, Thekla Haret, Adrian Kalavasis, Alkis Kóczy, László Lee, Barton
Mathioudaki, Angeliki McKay, Michael Olckers, Matthew Panageas, Ioannis
Papasotiropoulos, Georgios Perrussel, Laurent
Rastegari, Baharak Rey, Simon Roberts, Fred Serafino, Paolo Spanjaard, Olivier Sun, Zhaohong Sziklai, Balázs R.
Terzopoulou, Zoi Voudouris, Alexandros Wilczynski, Anaëlle Yang, Yongjie viii Organization
Abstracts of Invited Talks
Machine Learning (for) Security: Lessons Learned and Future Challenges
Battista Biggio1,2
1University of Cagliari, Italy
2Pluribus One
Abstract.In this talk, I will briefly review some recent advancements in the area of machine learning security [2] with a critical focus on the main factors which are hindering progress in thisfield. These include the lack of an underlying, systematic and scalable framework to properly evaluate machine-learning models under adversarial and out-of-distribution scenarios, along with suitable tools for easing their debugging. The latter may be helpful to unveilflaws in the evaluation process [7], as well as the presence of potential dataset biases and spurious features learned during training. I willfinally report concrete examples of what our laboratory has been recently working on to enable a first step towards overcoming these limitations [1, 3], in the context of Android [6] and Windows malware detection [4, 5].
Keywords: Machine learning Computer security Adversarial machine learning Malware detection
References
1. Biggio, B., et al.: Evasion attacks against machine learning at test time. In: Blockeel H., Kersting K., Nijssen S.,Železný F. (eds) ECML PKDD 2013. LNCS, vol. 8190, pp. 387–402.
Springer, Heidelberg (2013).https://doi.org/10.1007/978-3-642-40994-3_25
2. Biggio, B., Roli, F.: Wild patterns: ten years after the rise of adversarial machine learning.
Pattern Recogn. 84, 317–331 (2018)
3. Biggio, B., Nelson, B., Laskov, P.: Poisoning attacks against support vector machines. In:
29th ICML. pp. 1807–1814. Omnipress (2012)
4. Demetrio, L., Biggio, B., Lagorio, G., Roli, F., Armando, A.: Functionality-preserving black-box optimization of adversarial windows malware. IEEE Transactions on Information Forensics and Security 16, 3469–3478 (2021)
5. Demetrio, L., Coull, S.E., Biggio, B., Lagorio, G., Armando, A., Roli, F.: Adversarial EXEmples: a survey and experimental evaluation of practical attacks on machine learning for Windows malware detection. ACM Trans. Priv. Secur. (2021)
6. Demontis, A., et al.: Yes, machine learning can be more secure! a case study on android malware detection. IEEE Trans. Dep. Sec. Comp. 16(4), 711–724 (2019).
7. Pintor, M., et al.: Indicators of attack failure: Debugging and improving optimization of adversarial examples. CoRR abs/2106.09947 (2021)
Mind the Gap: Fair Division With Separation Constraints
Edith Elkind University of Oxford, UK [email protected]
Abstract.This is the extended abstract for the ADT’21 invited talk. It is based on a series of papers with Erel Segal-Halevi and Warut Suksompong [1, 2, 3].
Keywords: Cake cutting Land division Maximin fair share
Motivated by the social distancing rules, we consider the task of fairly distributing a divisible good among several agents under the additional assumption that every two agents’ shares must be separated. We start by analyzing the case where the good is the [0, 1] segment (usually referred to as‘cake’). In this model, the separation constraint is captured by specifying a separation parameter s such that for every pair of agents i, j and every pair of points x, y such that x is allocated to i and y is allocated to j it holds that |x– y| s; metaphorically, the cake is cut by a blunt knife of width s.
We observe that in this setting a proportional allocation cannot be guaranteed. We therefore focus on the solution concept of maximin fair share. Intuitively, wefirst ask each agent to determine their fair share by executing a mental experiment where they need to cut the cake into n s-separated pieces (where n is the number of agents), and are allocated the piece that they value the least; their fair share is then defined as the most they can guarantee for themselves under this protocol, and an allocation is considered fair if it provides each agent with a piece that they value at least as much as their fair share.
We show that a natural moving-knife protocol guarantees that each agent receives their fair share, i.e., maximin fair share allocations exist. However, to execute that protocol, agents need to be able to compute their fair shares, and it turns out that there is no finite algorithm that can accomplish this task. We circumvent this issue by providing an algorithm that approximates the agents’ shares up to an arbitrarily small error, as well as a polynomial-time algorithm for the case where all agents have piecewise constant valuations that are specified explicitly as part of the input.
We then extend our analysis of fair division with separation to richer settings: we consider fair division of a pie (i.e., a circular cake), land (i.e, a 2-dimensional good), and graphical cake (i.e., the ‘cake’ formed by edges of a graph). In many of these settings, maximin fair allocation is no longer guaranteed to exist. We therefore consider its ordinal approximation, defined as follows. Recall that, in the definition of maximin fair share, we asked each agent to perform a mental experiment where they cut the good into n pieces. We now ask each agent to re-run that experiment, but cut the good into k > n pieces, for a given value of k; we refer to the outcome of that experiment as k-fair share. Increasing the value of k corresponds to the agents being less ambitious in
terms of what they want to receive, so we are interested in the smallest value of k such that each agent can be guaranteed their k-fair share.
It turns out that for the circular cake with separation constraints it suffices to set k = n + 1; however, the circular cake is more challenging than the interval cake from an algorithmic perspective. For general graphical cakes, under a mild technical assump- tion, we can set k = n + f, where f is the feedback vertex set number of the underlying graph; in particular, if the graph is a tree, each agent can be guaranteed her maximin fair share (however, somewhat surprisingly, a natural extension of the moving knife protocol to trees may fail tofind a maximin fair allocation).
For land division, our results depend on the geometric shape of the land itself as well as the shapes of the agents’ pieces. For instance, if each agent has to be allocated a square piece of land, we can set k = 4n− 5. However, if agents’ pieces can be arbitrary axis-aligned rectangles (and the land itself is an axis-aligned rectangle), we get a much weaker upper bound of k = 2n+2, and converting it into afinite algorithm comes at an additional cost.
References
1. Elkind, E., Segal-Halevi, E., Suksompong, W.: Mind the gap: cake cutting with separation. In:
Proceedings of AAAI 2021
2. Elkind, E., Segal-Halevi, E., Suksompong, W.: Graphical cake cutting via maximin share. In:
Proceedings of IJCAI 2021
3. Elkind, E., Segal-Halevi, E., Suksompong, W.: Keep your distance: land division with sep- aration. In: Proceedings of IJCAI 2021
Mind the Gap: Fair Division With Separation Constraints xiii
Hierarchical Decision Models with Interacting Criteria: Preference Learning, Identifiability
and Explainability
Christophe Labreuche1,2
1Thales Research and Technology, Palaiseau, France [email protected]
2SINCLAIR AI Lab, Palaiseau, France
Multi-Criteria Decision Aiding (MCDA) aims at comparing several alternatives on the basis of multiple and possibly conflicting criteria. In industrial applications such as Air Traffic Management, elaborate decision models need to be used. Firstly, criteria are not independent as there are often statistical correlations and/or preferential dependencies among attributes. The leading model for capturing such interactions is the Choquet integral. Secondly, the number of criteria can be relatively large. In order to have an interpretable model, the set of criteria is organized in a hierarchical way instead of in a flat way. This means that criteria are aggregated by means of several nested aggregation functions. Each node in this hierarchy has a clear meaning to the decision maker. This allows enriching the representation power of the model while reducing the complexity thanks to a smaller number of parameters.
The combine use of hierarchical models and the representation of interacting cri- teria is expected to bring significant added values in real applications. We will expose recent results in three directions.
– Preference Learning. The existing approaches in Operation Research consists in eliciting each aggregation function and each marginal utility function separately from an interaction with the decision maker. However, providing only local pref- erence information may yield global inconsistency. The promise of preference learning is to go beyond these limitations, replacing time-consuming interactions with a user, by machine learning from a large quantity of (possibly noisy) prefer- ence data. The objective is to learn all parameters of the model simultaneously, which is challenging as the underlying optimization problem is no more convex.
We will present a Preference Learning approach based on the representation of the decision model as a neural network.
– Identifiability. To ease learning and interpreting the parameters of the model, there should not be two hierarchical models with different parameters and possibly dif- ferent hierarchies yielding the same decision function pointwise. We show identi- fiability for the hierarchical Choquet integral model. We are in particular able to relate structuring elements on the behavior of the model to the underlying hierarchy.
– Explainability. It is important in practice to explain the recommendations made by the model. Most of the time, the user does not need an in-depth explanation of the internal model mechanism, but he only wishes to understand which are the nodes in
the tree at the origin of the model outcome. This is obtained by computing an index measuring the influence of each node in the tree, on the preference between two alternatives. There are many connections with Feature Attribution (FA) in Machine Learning. Computing the level of contribution of a feature in a classification black-box model or that of a criterion in a MCDA model is indeed similar. The Shapley value is one of the leading concepts for FA. Unlike our situation, feature attribution only computes the influence of leaves in a model. We will show that the Shapley value is not appropriate on trees, when we are interested in knowing the contribution level of not only the leaves but also other nodes. We will then define a consistent value for trees.
References
1. Bresson, R., Cohen, J., Hüllermeier, E., Labreuche, C., Sebag, M.: Neural representation and learning of hierarchical 2-additive Choquet integrals. In: Proceedings of the Twenty-Eight International Joint Conference on Artificial Intelligence (IJCAI 2020), pp. 1984–1991, Yokohoma, Japan (2020)
2. Bresson, R., Cohen, J., Hüllermeier, E., Labreuche, C., Sebag, M.: On the identifiability of hierarchical decision models. In: Proceedings of the 18th International Conference on Principles of Knowledge Representation and Reasoning (KR 2021). Accepted 2021 3. Choquet, G.: Theory of capacities. Annales de l’Institut Fourier, 5, 131–295 (1953) 4. Fallah Tehrani, A., Labreuche, C., Hüllermeier, E.: Choquistic Utilitaristic Regressio. In:
Decision Aid to Preference Learning (DA2PL 2014) Workshop, Chatenay-Malabry, France, November 2014
5. Figueira, J., Greco, S., Ehrgott, M., (eds.) Multiple Criteria Decision Analysis: State of the Art Surveys 2nd edn. Kluwer Acad. Publ. (2016)
6. Fürnkranz, J., Hüllermeier, E.: Preference Learning. Springer-Verlag. Heidelberg.https://doi.
org/10.1007/978-3-642-14125-6(2010)
7. Grabisch, M., Labreuche, C.: A decade of application of the Choquet and Sugeno integrals in multi-criteria decision aid. Ann. Oper. Res. 175, 247–286 (2010)
8. Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F., Pedreschi. D.: A survey of methods for explaining black box models. ACM Comput. Surv. 51(6), Article 93 (2018) 9. Labreuche, C., Fossier, S.: Explaining multi-criteria decision aiding models with an extended
shapley value. In: Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI 2018), pp. 331–339, Stockholm, Sweden, July 2018 10. Labreuche, C.: A general framework for explaining the results of a multi-attribute preference
model. Artif. Intell. 175, 1410–1448, (2011)
11. Labreuche, C., Hüllermeier, E., Vojtas, P., Fallah Tehrani, E.: On the Identifiability of Models in Multi-Criteria Preference Learning. In: Decision Aid to Preference Learning (DA2PL 2016) workshop, Paderborn, Germany, November 2016
Hierarchical Decision Models with Interacting Criteria xv
12. Labreuche, C.: Explaining hierarchical multi-linear models. In: Proceedings of the 13th international conference on Scalable Uncertainty Management (SUM 2019), Compiègne, France, December 2019
13. Labreuche, C.: Explanation with the winter value: efficient computation for hierarchical Choquet integrals. In: Proceedings of the Sixteenth European Conference on Symbolic and Quantitative Approaches to Reasoning with Uncertainty (ECSQARU 2021). Accepted 2021 14. Lundberg, S., Lee, S.I.: Unified approach to interpreting model predictions. In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R., (eds.) 31st Conference on Neural Information Processing Systems (NIPS 2017), pp. 4768–4777, Long Beach, CA, USA (2017)
15. Shapley, L.S.: A value for n-person games. In: Kuhn, H.W., Tucker, A.W. (eds.) Contri- butions to the Theory of Games, vol. II, number 28 in Annals of Mathematics Studies, pp. 307–317. Princeton University Press (1953)
xvi C. Labreuche
Contents
Computational Social Choice and Preference Modelling
Aggregating Preferences Represented by Conditional Preference Networks . . . 3 Abu Mohammad Hammad Ali, Howard J. Hamilton, Elizabeth Rayner,
Boting Yang, and Sandra Zilles
Measuring Nearly Single-Peakedness of an Electorate:
Some New Insights . . . . 19 Bruno Escoffier, Olivier Spanjaard, and Magdaléna Tydrichová
Preference Aggregation in the Generalised Unavailable Candidate Model . . . . 35 Arnaud Grivet Sébert, Nicolas Maudet, Patrice Perny,
and Paolo Viappiani
Simultaneous Elicitation of Scoring Rule and Agent Preferences for Robust
Winner Determination . . . . 51 Beatrice Napolitano, Olivier Cailloux, and Paolo Viappiani
Preference Elicitation
Incremental Elicitation of Preferences: Optimist or Pessimist? . . . . 71 Loïc Adam and Sébastien Destercke
Probabilistic Lexicographic Preference Trees . . . . 86 Xudong Liu and Miroslaw Truszczynski
Incremental Preference Elicitation with Bipolar Choquet Integrals . . . . 101 Hugo Martin and Patrice Perny
Preference Aggregation and Voting
In the Beginning There Were n Agents: Founding and Amending
a Constitution . . . . 119 Ben Abramowitz, Ehud Shapiro, and Nimrod Talmon
Unveiling the Truth in Liquid Democracy with Misinformed Voters . . . . 132 Ruben Becker, Gianlorenzo D’Angelo, Esmaeil Delfaraz,
and Hugo Gilbert
Computing Kemeny Rankings from d-Euclidean Preferences . . . . 147 Thekla Hamm, Martin Lackner, and Anna Rapberger
Iterative Deliberation via Metric Aggregation . . . . 162 Gil Ben Zvi, Eyal Leizerovich, and Nimrod Talmon
Manipulation in Voting
Obvious Manipulability of Voting Rules . . . . 179 Haris Aziz and Alexander Lam
Manipulation in Communication Structures of Graph-Restricted Weighted
Voting Games . . . . 194 Joanna Kaczmarek and Jörg Rothe
Strategic Voting in Negotiating Teams. . . . 209 Leora Schmerler and Noam Hazon
The Nonmanipulative Vote-Deficits of Voting Rules . . . . 224 Yongjie Yang
Fair Division and Resource Allocation
Allocating Indivisible Items with Minimum Dissatisfaction
on Preference Graphs. . . . 243 Nina Chiarelli, Clément Dallard, Andreas Darmann, Stefan Lendl,
Martin Milanič, Peter Muršič, Nevena Pivač, and Ulrich Pferschy
On Fairness via Picking Sequences in Allocation of Indivisible Goods. . . . 258 Laurent Gourvès, Julien Lesca, and Anaëlle Wilczynski
On Reachable Assignments in Cycles . . . . 273 Luis Müller and Matthias Bentert
Minimizing and Balancing Envy Among Agents Using Ordered
Weighted Average. . . . 289 Parham Shams, Aurélie Beynier, Sylvain Bouveret, and Nicolas Maudet
Algorithmic Decision Theory
Interactive Optimization of Submodular Functions Under Matroid
Constraints . . . . 307 Nawal Benabbou, Cassandre Leroy, Thibaut Lust, and Patrice Perny
Necessary and Possible Interaction in a 2-Maxitive Sugeno
Integral Model . . . . 323 Paul Alain Kaldjob Kaldjob, Brice Mayag, and Denis Bouyssou
xviii Contents
Coalition Formation
Democratic Forking: Choosing Sides with Social Choice . . . . 341 Ben Abramowitz, Edith Elkind, Davide Grossi, Ehud Shapiro,
and Nimrod Talmon
Hedonic Diversity Games Revisited. . . . 357 Andreas Darmann
Stable Matchings
Multi-agent Reinforcement Learning for Decentralized Stable Matching . . . . . 375 Kshitija Taywade, Judy Goldsmith, and Brent Harrison
Lazy Gale-Shapley for Many-to-One Matching with Partial Information. . . . . 390 Taiki Todo, Ryoji Wada, Kentaro Yahiro, and Makoto Yokoo
Participatory Budgeting
Participatory Funding Coordination: Model, Axioms and Rules. . . . 409 Haris Aziz and Aditya Ganguly
Complexity of Manipulative Interference in Participatory Budgeting . . . . 424 Dorothea Baumeister, Linus Boes, and Johanna Hillebrand
Author Index . . . . 441 Contents xix