Optimized Score Transformation for Fair Classification
[edit]
Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, PMLR 108:16731683, 2020.
Abstract
This paper considers fair probabilistic classification where the outputs of primary interest are predicted probabilities, commonly referred to as scores. We formulate the problem of transforming scores to satisfy fairness constraints while minimizing the loss in utility. The formulation can be applied either to postprocess classifier outputs or to preprocess training data, thus allowing maximum freedom in selecting a classification algorithm. We derive a closedform expression for the optimal transformed scores and a convex optimization problem for the transformation parameters. In the population limit, the transformed score function is the fairnessconstrained minimizer of crossentropy with respect to the optimal unconstrained scores. In the finite sample setting, we propose to approach this solution using a combination of standard probabilistic classifiers and ADMM. Comprehensive experiments comparing to 10 existing methods show that the proposed FairScoreTransformer has advantages for scorebased metrics such as Brier score and AUC while remaining competitive for binary labelbased metrics such as accuracy.
Related Material


