Fair classification using mutual information and kernel density estimation

Abstract

As machine learning increasingly affects decisions in domains protected by anti-discrimination laws, one major criterion in the design of machine learning classifiers is to ensure fairness: guaranteeing the irrelevancy of a prediction to sensitive attributes such as race and gender. This paper presents two fair classifiers: one is based on a prominent information-theoretic notion, mutual information

the other employs a well-known statistical approach, kernel density estimation (KDE). One feature of our MI-based approach is that it provides a theoretical interpretation of prior fair classifiers that rely on an adversarial learning framework. On the other hand, our KDE-based fair classifier has a great training stability that many adversarial learning approaches are not equipped with. We focus on the two well-known group fairness measures, demographic parity and equalized odds. We conduct extensive experiments both on synthetic and benchmark real datasets to demonstrate that our approaches outperform state of the arts in accuracy-fairness tradeoff.