Fair and Explainable Collective Decision

Local organizing committee

• Chloé Le Bail (LISN, Université Paris-Saclay)
• Vincent Mousseau (MICS, CentraleSupélec, Université Paris-Saclay)
• Wassila Ouerdane (MICS, CentraleSupélec, Université Paris-Saclay)
• Fabien Tarissan (CNRS, ENS Paris-Saclay)
• Anaëlle Wilczynski (MICS, CentraleSupélec, Université Paris-Saclay)

Fair and Explainable Collective Decision

Many real-life problems in our societies involve making a collective decision, where a group of individuals needs to aggregate their individual preferences over different alternatives to make a decision for the group. As concrete real-world examples, one can cite the choice of projects to fund in cities, known as participatory budgeting, a participatory democratic approach which is increasingly adopted in many countries around the world [Cabannes, 2004, Aziz and Shah, 2020], or the assignment of students to courses or universities (see, e.g., in France the Parcoursup application for graduate studies), which falls into the class of matching problems under preferences [Klaus et al., 2016], where agents need to be matched to elements (tasks, roommates, jobs, etc.) based on their preferences.

With the increasing use of algorithms and AI tools in systems governing our life choices (job recruitment, insurance covering, university assignment), important decisions for the agents can be made in systems assisting collective decision. Therefore, to ensure confidence and participation
in the system, it is crucial to guarantee that algorithms used for computing collective decision are fair to the agents. A fair decision should respect the expressed preferences of the agents and should not discriminate any population. In participatory budgeting, the decision has a collective
implementation, in the sense that all agents will be impacted by the decision in the same way: all citizens will benefit from the same selected projects. Under this perspective, fairness imposes to adequately represent all agents’ preferences in the outcome, e.g., in a proportional way [Aziz et al.,
2018, Peters et al., 2021, Los et al., 2022]. In contrast, in matching problems, the decision implies different treatments for different agents since they may not be assigned to the same elements. In consequence, in this setting, the question of fairness is more oriented towards the agents’ individual
satisfaction and feeling of equal treatment.

The guarantee of fairness in collective decision making can be achieved via the justification of the final decision to satisfy a given appropriate fairness concept. This involves designing realistic and achievable fairness concepts, but also to explain that the final outcome is actually fair. Justifying that a given solution is fair is related to the ability of explaining decisions. Explainable Artificial Intelligence (XAI) is a hot topic in AI [Barredo Arrieta et al., 2020], which has even become a political and legal concern (see, e.g., the Villani report on AI in 2018 [Villani et al., 2018] and the European General Data Protection Regulation (GDPR) [Goodman and Flaxman, 2017]). The goal is to explain decisions made by systems based on algorithms, which may appear as black boxes, and making them more transparent. Concerning collective decision making, researchers agree that transparency, explainability and justificability of collective decision processes are the new challenges
in social choice [Nunez, 2019, Procaccia, 2019].

In computational social choice [Brandt et al., 2016], there is a long tradition of axiomatic characterizations of preference aggregation methods [Arrow et al., 2002, 2011, Procaccia, 2019], which can be seen as rule-based explanations. One can indeed argue that a given solution is a good choice if it is the only one satisfying a given combination of desirable properties. To this aim, we need to define axioms which capture the merits of a solution. It remains to be seen whether the chosen combination of axioms is possible. This can notably be checked by using computer-aided techniques to derive impossibility theorems [Geist and Peters, 2017]. But, such computer-aided methods can also be used in order to derive justification of the outcomes of voting systems. As defined by Boixel and Endriss [2020], a valid justification for a given collective decision is the combination of a set of normative principles (axioms) and the explanation that the outcome actually follows them. The initial studies in this line of research focused on the generation of justification algorithms for very specific scenarios in voting [Cailloux and Endriss, 2016]. More generality could then be reached for generating justification by using more optimized AI tools such as constraint programming [Boixel and Endriss, 2020] or SAT solving [Boixel et al., 2022]. Hence, important research questions remain, on the design of adequate notions of fairness and their explanation. Moreover, this raises key legal issues that need to be addressed, and have been largely under-investigated.

The goal of this workshop is to bring together researchers from different communities in order to share insights on these topics and progress towards better understanding the requirement of fairness and its explanation in collective decision making, especially on the two concrete topics of participatory budgeting and matchings under preferences. The main research questions we can investigate in these two settings are the following:

References

Schedule

The workshop will take place over two weeks from March 30th to April 10th at Institut Pascal (Orsay, France).

The global schedule of the two weeks is composed of:

• a global introduction and conclusion for the workshop,

• a description of GDR RADIA and a GDR invited talk,

• two transverse invited talks,

• four slots for participants’ short presentations,

• a doctoral consortium,

• a social event,

• two social dinners,

• two working blocks dedicated to specific topics:

  • Week A: participatory budgeting 

  • Week B: matching under preferences

Each working block is composed of 11 sessions of 1h30 structured as follows:

• one tutorial, given by an expert of the topic,

• a global discussion in order to identify interesting research questions and to form subgroups,

• two invited talks,

• five working sessions in subgroups and a last one dedicated to the writing of the report and preparation of the presentation of the subgroup,

• one final session where each subgroup presents its work to the whole audience.

The global preliminary organization of the two weeks is presented below.

Global schedule of the two weeks