Training individually fair ML models with Sensitive Subspace Robustness

TL;DR

This paper introduces a distributionally robust optimization method to train machine learning models that maintain fairness by being invariant to sensitive input perturbations, addressing biases related to gender and ethnicity.

Contribution

It formalizes a new fairness criterion based on sensitive subspace robustness and proposes an optimization approach to enforce it during training.

Findings

Effective in reducing gender and racial biases in ML tasks

Improves fairness without sacrificing overall performance

Demonstrates robustness against sensitive attribute perturbations

Abstract

We consider training machine learning models that are fair in the sense that their performance is invariant under certain sensitive perturbations to the inputs. For example, the performance of a resume screening system should be invariant under changes to the gender and/or ethnicity of the applicant. We formalize this notion of algorithmic fairness as a variant of individual fairness and develop a distributionally robust optimization approach to enforce it during training. We also demonstrate the effectiveness of the approach on two ML tasks that are susceptible to gender and racial biases.