Learning Non-Discriminatory Predictors

TL;DR

This paper investigates methods for learning predictors that satisfy equalized odds fairness constraints, highlighting limitations of post-hoc corrections, proposing a nearly-optimal statistical approach, and introducing a tractable relaxation.

Contribution

It introduces a nearly-optimal statistical procedure for fair learning, analyzes the intractability of existing methods, and proposes a tractable relaxation of the fairness criterion.

Findings

Post-hoc correction can be highly suboptimal.

Proposes a nearly-optimal statistical learning procedure.

Identifies intractability in the computational problem.

Abstract

We consider learning a predictor which is non-discriminatory with respect to a "protected attribute" according to the notion of "equalized odds" proposed by Hardt et al. [2016]. We study the problem of learning such a non-discriminatory predictor from a finite training set, both statistically and computationally. We show that a post-hoc correction approach, as suggested by Hardt et al, can be highly suboptimal, present a nearly-optimal statistical procedure, argue that the associated computational problem is intractable, and suggest a second moment relaxation of the non-discrimination definition for which learning is tractable.