Abstract

The most prevalent notions of fairness in machine learning are statistical and coarse: they fix a small collection of pre-defined groups or attributes (such as race or gender), and then ask for parity of some statistic of the classifier (such as false positive rate) across these groups. Constraints of this form are susceptible to intentional or inadvertent “fairness gerrymandering”, in which a classifier appears to be fair on each individual group, but badly violates fairness on one or more subgroups defined over the protected attributes.

We propose to instead demand notions of fairness across exponentially (or infinitely) many subgroups, defined by a structured class of functions over the protected attributes. This interpolates between statistical definitions of fairness and recently proposed individual notions of fairness. While the problem of auditing a given classifier for subgroup fairness can be computationally intractable in the worst case, we show that it can itself be cast as an instance of weighted classification, and thus standard learning algorithms can be applied.

We then present algorithms that provably converge to the best subgroup-fair classifier. The algorithms are based on a formulation of subgroup fairness as a two-player zero-sum game between a Learner and an Auditor. We present extensive empirical validation on a number of datasets in which fairness is a concern, and demonstrate that appealing trade-offs between accuracy and subgroup fairness are possible in practice.

Joint work with Seth Neel, Aaron Roth, and Zhiwei Steven Wu.