On Fairness of Classification in Machine Learning

Abstract

As our society becomes increasingly automated, there is a social concern for algorithmic decision-making to be fair and objective. In this thesis, we initiate the study with an overview of several criteria for group fairness, their limitations and motivations, and the criterion of individual fairness. We start with the case of a single-classifier, and extend the fairness properties to systems using multiple classifiers in composition. We demonstrate how to construct such systems, and find that fairness in social situations varies greatly with context. We find that that classifiers that are fair-in-isolation may not necessary yield fair systems in naive composition, and fair systems can be constructed from individually unfair classifiers. Finally, we examine the behavior of group fairness criteria under systems of multiple classifiers.