SUMMARY

This project is positioned in the core of the emerging research area on social influence analysis but goes further in trying to demonstrate not only that this framework can be applied to other research domains through a property-driven approach, but also that the algorithmic and strategic aspects play an important role in the design and selection of specific social ranking solutions. Unlike previous generations of influence measurement methods for social structures, this project brings insights based on different techniques from artificial intelligence including multi-agent systems, compact representation, algorithmic game theory, computational social choice, and social network analysis. Our main objective is to show that the portfolio of models proposed under the umbrella of a qualitative theory for social influence analysis is more adapted to answer important questions arising from different domains of collective decision making like voting, social networks, argumentation, and MCDA.

Problem

What is the common point between the affectation of students to universities (in particular, in France, the Parcoursup algorithm), the influence of someone in a social network (like Twitter), the responsibility of a formula in the inconsistency of a belief base, the impact and synergy of some criteria in a multi-criteria decision making situation? In all these situations we are confronting with a set of objects/persons on which we have information on the results of some coalitions/groups, and where it is interesting (if not decisive) to obtain information on individuals. An idealised version of this problem has been studied for long in (cooperative) game theory, especially under the notion of power index. But in real situations, like the ones given above, we are far from this idealised setting. In particular we often lack of information, and in a lot of applications we only have ordinal information (what coalition performs better, or is preferable, to another, etc.). But we also face other typical problems in artificial intelligence, like the representation of the information, some constraints (restrictions), the computational complexity, and the concrete implementation. So this important problem, requires to melt together, in an innovative way, tools and expertise from Artificial Intelligence, Game Theory and Social Choice Theory.

Objectives

In this project we want to provide a more flexible theory of cooperative interaction situations and power indices based on the evidence that the nature of available information about the interaction of individuals and groups is mostly ordinal. Our main objective is to show that the portfolio of models and solutions proposed under the umbrella of an ordinal theory is more adapted to answer important questions arising from different domains of collective decision making like explicability end interpretability in artificial intelligence, social network formation, belief merging and multi-criteria decision analysis (MCDA).

These issues will be addressed by researchers whose expertise includes a range over the following domains in artificial intelligence across computer science and economics:

- Cooperative game theory: for the re-formulation of classical concepts of solutions for cooperative games in our ordinal framework and the analysis of models of interaction restrictions according to our interpretation of ordinal comparison of coalitions; and for the analysis of dynamics of coalition formation and the related algorithmic and complexity issues.

- Social choice: for the problem related to the axiomatic design of collective decision making mechanisms of social ranking.

- Computational social choice: for the impact of computational complexity of social ranking solution to different forms of strategic behaviour.

- Compact representation: for the definition of a general representation of our ordinal coalitional framework that can be used as a prototyping tool for studying new interaction situations and finding efficient algorithms that go with it.

- Explainable AI: for assessing consistent and accurate measures of features' importance for the final prediction of classification models and help human experts to understand the results;

- Influence measurement in networks: for measuring the centrality of agents over a social network with respect to their ability to influence alternative groups of followers' choice of action in certain instances.

The originality of our project is summarized by the following list of innovative tasks:

- to conceive a novel ordinal theory of cooperative games for measuring power and influence in coalitional situations;

- to focus on a property-driven design of social ranking solutions that will be used as a top-down approach aimed at splitting the complex behaviour of groups or coalitions into more intelligible interaction situations;

- to explore natural applications of methods for compact preference representation to social ranking computation and to coalition formation;

- to formulate a portfolio of solutions accompanied with a road-map of their properties to drive users in their practice;

- to explain the ordinal influence of criteria in Parcoursup, the national admissions platform for the first year of higher education.

- to implement most of the solutions proposed in this project as a software package.

These original issues, as well as other problems described, do not correspond necessarily to one single task, but more to the interaction between different subtasks allocated over different work-packages (WPs).