Exposé : Algorithmic Fairness Through Counterfactual Analysis and Optimal Transport
Mardi 18 février se la deuxième journée de l’année universitaire 2024–2025 du séminaire interne du CRM-CNRS.
Je présenterai les thématiques sur lesquelles nous travaillons avec Arthur Charpentier et Agathe Fernandes Machado, portant sur l’équité algorithmique (algorithmic fairness).
This presentation will explore ways to address fairness issues in predictive modeling, echoing societal or legal demands. It will begin by examining how different disciplines define discrimination and fairness, setting the stage for a more focused discussion on algorithmic fairness. Fairness will be specifically defined within the insurance industry, considering its impact on decision-making processes, with a focus on predictive models that return risk scores. The presentation will then introduce methods for quantifying the (un)fairness of predictive models. Finally, the talk will present a mitigation approach based on counterfactual analysis. More precisely, by assuming a causal structure represented by Directed Acyclic Graphs (DAGs), a sequential method for constructing counterfactuals will be proposed, based on optimal transport. The presentation will also address the often-overlooked case of categorical data, where data is represented in the simplex, and Dirichlet transport is applied to ensure fairness between protected and non-protected groups.
Les diapositives (dont une partie importante a été, à nouveau, empruntée à celles d’Arthur) sont en ligne: Diapositives