Impact of Data Processing on Fairness in Supervised Learning

We study the impact of pre and postprocessing for reducing discrimination in data-driven decision makers. We first analyze the fundamental trade-off between fairness and accuracy in a preprocessing approach, and propose a design for a preprocessing module based on a convex optimization program, which can be added before the original classifier. This leads to a fundamental lower bound on attainable discrimination, given any acceptable distortion in the outcome. Furthermore, we reformulate an existing postprocessing method in terms of our accuracy and fairness measures, which allows comparing postprocessing and preprocessing approaches. We show that under some mild conditions, preprocessing outperforms postprocessing. Finally, we show that by the appropriate choice of the discrimination measure, the optimization problem for both pre and postprocessing approaches will reduce to a linear program and hence can be solved efficiently.