Fair Logistic Regression via Penalized Mean-Field Variational Inference

Abstract

Fairness in Machine Learning (ML) has recently risen in importance due to its demand by both society and institutions, either private or public. This is due to the discrimination or disparities that may arise as a consequence of biases in the data sets that are used for training the prediction algorithms. As a consequence, the applicability of such algorithms may be limited. In this paper we propose a fair version of the logistic regression via Mean-Field Variational Bayes (MFVB). Fairness is defined in terms of a carefully selected metric that, not only is interpretable for the considered logistic regression problem, but also leads to tractable MFVB equations. In addition, it allows for the problem to be solved in a private manner, further increasing the applicability of the method. The MFVB equations are solved through a Coordinate Ascent Variational Inference (CAVI) algorithm, where a proper selection of the hyper-parameters is shown to be crucial to assure accurate classification rates. Numerical experiments illustrate how the novel approach provides a trade-off between the accuracy of the method and the predictions’ level of fairness that eases the decision making process.