61
5

Metric-Fair Classifier Derandomization

Abstract

We study the problem of classifier derandomization in machine learning: given a stochastic binary classifier f:X[0,1]f: X \to [0,1], sample a deterministic classifier f^:X{0,1}\hat{f}: X \to \{0,1\} that approximates the output of ff in aggregate over any data distribution. Recent work revealed how to efficiently derandomize a stochastic classifier with strong output approximation guarantees, but at the cost of individual fairness -- that is, if ff treated similar inputs similarly, f^\hat{f} did not. In this paper, we initiate a systematic study of classifier derandomization with metric fairness guarantees. We show that the prior derandomization approach is almost maximally metric-unfair, and that a simple ``random threshold'' derandomization achieves optimal fairness preservation but with weaker output approximation. We then devise a derandomization procedure that provides an appealing tradeoff between these two: if ff is α\alpha-metric fair according to a metric dd with a locality-sensitive hash (LSH) family, then our derandomized f^\hat{f} is, with high probability, O(α)O(\alpha)-metric fair and a close approximation of ff. We also prove generic results applicable to all (fair and unfair) classifier derandomization procedures, including a bias-variance decomposition and reductions between various notions of metric fairness.

View on arXiv
Comments on this paper