On the Generalizability of Fair Representations

Abstract

The concept of a fair representation was introduced more than a decade ago and has inspired many follow-up works with the promising idea of a transferable representation that guarantees fairness for many downstream tasks. However, it has been shown that this promise of guaranteed fairness is impossible in its proposed generality. That is, impossible without further restricting how and for which tasks the representation is being used. In this work, we consider a more attainable case in which classifiers are restricted, either structurally or through meeting certain accuracy-requirements (or both). We analyze whether the above impossibility results then still hold and how they could be overcome.