A fair classifier using kernel density estimation

Abstract

As machine learning becomes prevalent in a widening array of sensitive applications such as job hiring and criminal justice, one critical aspect in the design of machine learning classifiers is to ensure fairness: Guaranteeing the irrelevancy of a prediction to sensitive attributes such as gender and race. This work develops a kernel density estimation (KDE) methodology to faithfully respect the fairness constraint while yielding a tractable optimization problem that comes with high accuracy-fairness tradeoff. One key feature of this approach is that the fairness measure quantified based on KDE can be expressed as a differentiable function w.r.t. model parameters, thereby enabling the use of prominent gradient descent to readily solve an interested optimization problem. This work focuses on classification tasks and two well-known measures of group fairness: demographic parity and equalized odds. We empirically show that our algorithm achieves greater or comparable performances against prior fair classifers in accuracy-fairness tradeoff as well as in training stability on both synthetic and benchmark real datasets.