SEMINAIRE LABO - François HU (ISFA et Milliman France)

Predictive models are increasingly used to score individuals in domains such as credit allocation, insurance pricing, hiring, and marketing. When a model yields different outcomes across groups, regulators, auditors, and decision-makers naturally ask two questions: how large is the disparity, and what drives it? Simply removing the sensitive attribute is generally insufficient, as correlated covariates may continue to encode the same information. The talk begins with an accessible introduction to these questions through the lens of the Oaxaca-Blinder decomposition, a classical tool in economics. While highly influential, this approach focuses solely on differences in average outcomes, even though two groups may receive the same average prediction while facing markedly different prediction distributions. I will then present a framework, published at AAAI 2026, that quantifies disparities between entire prediction distributions using optimal transport, decomposes them into direct and indirect effects, and characterizes in closed form the most accurate linear predictor satisfying a prescribed fairness constraint. Time permitting, I will also discuss recent work on exact feature-level attribution of these disparities, together with uncertainty quantification through confidence intervals, illustrated on applications in crime and healthcare expenditure data.

Salle 2303, 14h à 15h.