Fairness in Machine Learning: Why and How?

Abstract

Services based on machine learning are increasingly present in our everyday lives. While such application make promises of its improvement, they also pose considerable risks if machine learning models do not perform as expected. One specific issue related to the quality of learnt models which has recently gained considerable visibility is their unfairness. Namely, it has been noted that the decisions of machine learning models sometimes reflect human biases against some historically discriminated groups of people, thus unintendedly perpetuating the discrimination. In this paper we discuss why is the fairness of machine learning models important, by revisiting some notable examples of discrimination committed by the models and discuss different notions of fairness. We discuss how to measure the fairness of such models and how to achieve it, reflecting on both algorithmic and non-technical aspects of this effort. We present several fairness ensuring methods representative of different fairness paradigms, one of them being our own.