From Parity to Preference-based Notions of Fairness in Classification

Many notions of fairness in data-driven decision making are inspired by the concept of discrimination in social sciences and law, and focus on ensuring parity (equality) in treatment or outcomes for different social groups. In this paper, we propose preference-based notions of fairness with the goals of avoiding potential ‘reverse-discrimination’ and enabling high decision accuracy. We introduce tractable proxies to design convex boundary-based classifiers that satisfy these new notions of fairness and show on the ProPublica COMPAS dataset that these notions allow for greater decision accuracy than parity-based fairness. This paper is a shortened version of arxiv:1707.00010.