System and method for mitigating bias in classification scores generated by machine learning models

Abstract

A method, system, and computer-readable medium are disclosed for detecting and mitigating bias in a trained machine learning model. The method includes the steps of: training the model based on a training data set; detecting bias in the model relative to a protected class; identifying one or more groups of input variables that contribute to the bias; and mitigating bias in the model. Mitigating the bias is performed by constructing a post-processed score function that either (a) neutralizes or partially neutralizes one or more groups of input variables in the input vector of the model, or (b) utilizes a fair score approximation of the model to project the distributions for the protected class and/or the unprotected class to substantially match. In an embodiment, detecting bias in the trained model is performed by comparing distribution for two or more subpopulations based on a distance metric, such as a Wasserstein distance.