EWG on Preference Handling

Advances in Preference Handling

Multidisciplinary Working Group affiliated to EURO

Quai de la Dorade Toulouse

The preparation of the next workshop on Advances in Preference Handling has started. The workshop will be announced in the next weeks.

Jan 24, 2022

INSIGHTS FROM ADT IN TOULOUSE

LinkedIn Group on Preference Handling
Mail to the Working Group on Advances in Preference Handling

Research is a long-term endeavor driven by questions and not by the means. Good intentions are not sufficient to do research. It is necessary that you really want to know the answer of a research question. To come up with groundbreaking results, it is important to be perseverant, to continue the research even when support is weak or missing, and to report negative results in the same way as positive ones. Negative results help us to understand the limits of an approach. Many of the modern sciences have developed around impossibility results explaining the failure of century-long quests like that of building a perpetual motion machine or turning lead into gold. Even logic and computing do not escape from impossibility results as we all know. Scientific research is thus based on the honest reporting of results, whether these are those that we hoped for or the complete opposite of them. And often these results are not the fruit of a short project, but that of a program spanning one or more decades.

The invited talks given at the Seventh International Conference on Algorithmic Decision Theory ADT 2021 provided some good examples of results that have been achieved thanks to longer-term research programs.

Christophe Labreuche (Thales Research & Technology, France) presented a comprehensive approach to preference handling for hierarchical Choquet integrals. He motivated the interest of this preference model by several real-world applications ranging from quality assessment of flight tracking in air-traffic control up to the assessment of passenger services offered by a metro line. While hierarchical Choquet integrals are quite new, their ingredients, the Choquet integrals, have been studied by Christophe and colleagues for many years as a means to model interactions between multiple criteria. Christophe discussed mathematical properties and gave conditions under which the parameter values of a hierarchical Choquet integral identify such a model in a unique way. He furthermore showed that hierarchical Choquet integrals can be learned by a deep neural network that reflects the structure of this integral. An important part of the talk has been dedicated to the explanations of preferences. These explanations determine the relative contribution of each criterion when comparing two alternatives. Christophe showed that the popular Shapley value is insufficient for this purpose when the utilities of the two alternatives are computed by a hierarchical Choquet integral. He proposed to use a value defined by Eyal Winter instead of the Shapley-value and developed an algorithm for computing explanations based on this Winter-value.

Edith Elkind (University of Oxford, UK) presented recent results on fair allocation of a divisible good among multiple agents according to their preferences. Edith introduced a variant of this problem where the shares allocated to different agents need to be separated by a minimum amount. She motivated these separation constraints by a couple of examples such as the social distancing of vendors on a fair as well as the allocation of land while keeping different plots separated in order to prevent cross-fertilization. Edith studied the impact of these constraints on solution concepts such as maximin fairness and on allocation algorithms that are based on the Robertson-Webb model for acquiring preference information from the agents. The main result is negative and states that no finite algorithm can compute the desired allocation. Edith explored approximation methods as well as conditions under which an exact polynomial allocation algorithm exists. Starting with the division of intervals, Edith then continued this study for the fair division of more complex goods such as pies, two-dimensional areas, and graphs. Whereas the whole study has been motivated by the pandemic and carried out during this period, it was helped by a good understanding of the basic fair allocation problem, which has been the result of a long-term research effort.

A third invited talk was dedicated to machine learning (ML) for security. Battista Biggio (University of Cagliari, Italy) presented the lessons learned from a decade-long research effort in this area. Whereas machine learning methods such as linear regression have been used in email spam filtering for a long time, they are vulnerable to different kinds of attacks. Examples are evasion attacks which consist in finding perturbations of inputs that flip the decision. Another example are poisoning attacks where the attacker sends data (such as spam emails) to compromise the learning algorithm. As a consequence, the spam filter will be less effective. In certain cases, a single wrongly labeled data point can impact the accuracy of the learning algorithm by 10%. Battista analyzed reasons for the vulnerability of ML methods and proposed a framework that protects the learning algorithm during training, conducts security testing during validation, and monitors the security during deployment. In order to defend an ML model against attacks, he proposes to evaluate machine-learning models while taking sophisticated adversarial attacks into account.

These three invited talks exemplify research results of high quality that require time and experience. You cannot simply accelerate research by increasing the staffing of a project. Research questions have their own dynamics and ideas need to be challenged and accepted by the scientific community before you can start a new idea that depends on previous research. It is difficult to shortcut those cycles.

ADT has been held as a hybrid conference on November 3-5, 2021 in Toulouse, France and it was great to meet in person after a long break of 21 months. A bit more than two third of the participants physically met onsite at the University of Toulouse 1 Capitole in a spacious auditorium while respecting all sanitary measures. Thanks to the local organizers Umberto Grandi, Sylvie Doutre, Laurent Perrussel, Pascale Zaraté, and Rachael Colley, the whole event went smoothly and it was a pleasure to participate. The program co-chairs David Rios Insua and Dimitris Fotakis had put an interesting program together including the three invited talks and twenty-seven regular talks. The talks covered topics such as preference elicitation and modeling, preference aggregation, computational social choice, voting, coalition formation, fair division and resource allocation, stable matchings, and participatory budgeting. The conference was preceded by a PhD day and thus allowed a large number of PhD students to participate in this event. The conference proceedings have been published as Volume 13023 of the Lecture Notes in Artificial Intelligence series.

For 2022, there are plans to organize an M-PREF workshop and a DA2PL meeting. The calls for papers will be announced on this web site as soon as they are available.

Among the conferences in 2022 that are related to preference handling, we would like to mention the following ones. The 36th AAAI Conference on Artificial Intelligence AAAI-22 will be held virtually on February 22-March 1, 2022. A virtual format has also been chosen for the International Conference on Autonomous Agents and Multi-Agent Systems AAMAS-2022, which will be held on May 9–13, 2022.

The next two conferences are planned as physical events. The 32nd EURO Conference will be held in Espoo, Finland on July 3-6, 2022 and has an abstract submission deadline of March 4, 2022. Shortly after, the 31st International Joint Conference on Artificial Intelligence and the 25th European Conference on Artificial Intelligence IJCAI-ECAI 2022 will be held in Vienna, Austria on July 23-29, 2022.

Finally, we would like to mention the Fifth AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society AIES-22 which will be held in Oxford on August 1-3, 2022 as a hybrid event. Its paper submission deadline is on February 22, 2022.

FEB 2022

MAY 2022

CALENDAR

AUG 2022