Stochastic Optimization with Fairness Constraints

Information Systems and Operations Management 

Intervenant: Daniel Kuhn (EPFL)

Salle Bernard Ramanantsoa

Abstract

The last decade has witnessed a surge of algorithms that have a consequential impact on our daily lives. Machine learning methods are increasingly used, for example, to decide whom to grant or deny loans, college admission, bail or parole. Even though it would be natural to expect that algorithms are free of prejudice, it turns out that cutting-edge AI techniques can learn or even amplify human biases and may thus be far from fair. Accordingly, a key challenge in automated decision-making is to ensure that individuals of different demographic groups have equal chances of securing beneficial outcomes. In this talk we first highlight the difficulties of defining fairness criteria, and we show that a naive use of popular fairness constraints can have undesired consequences. We then characterize situations in which fairness constraints or unfairness penalties have a regularizing effect and may thus improve out-of-sample performance. We also identify a class of unfairness-measures that are susceptible to efficient stochastic gradient descent algorithms, and we propose a statistical hypothesis test for fairness.