Learning Smooth And Fair Representations

This paper explores the statistical properties of fair representation learning, a preprocessing method that preemptively removes the correlations between features and sensitive attributes by mapping features to a fair representation space. The demographic parity of a representation can be certified from a finite sample if and only if the chi-squared mutual information between features and representations is finite for all features distributions. Empirically, we find that smoothing representations provides generalization guarantees of fairness certificates, which improves upon existing fair representation learning approaches. On four datasets we simulate many downstream users and show that our approach, AGWN, is the only one that generates representations whose fairness properties are robust to many downstream users.