Evaluating the Reductions Intervention for Enforcing Fairness in Classi cation

Abstract

Fairness is a topic in machine learning con- cerned with removing biases present in machine learning systems. One method for removing such biases is the reduc- tions approach by Agarwal et al.1, which has the notable advantage of being usable for a wide range of fairness de nitions. Past work shows that this method gives a good fairness-accuracy tradeo compared to other methods that enforce fairness. In this paper, the performance of this method will be further investigated. To do this, I will compare the method against a preprocessing method for the demo- graphic parity fairness defi nition.