R. Jiménez Llamas, E. Carrizosa Priego, P. Ramirez Cobo
Fairness in machine learning (ML) is a recent and prominent research area with the aim of correcting discriminatory solutions in prediction algorithms caused by biases in the training dataset. In this work we present a novel method for fairness in logistic regression based on variational inference and the Mean Field approximation. Specifically, we consider a penalization term, proportional to the unfairness degree of the solution, that is chosen in a way that allows for a simple modification of the well-known CAVI iterative method. As a result, the computational cost of the algorithm does not increase significantly for high dimensions. Additionally, the method provides a clear tradeoff between the accuracy of the prediction model and the fairness degree of the solution, a fact that facilitates decision-making. The novel method shall be illustrated using both simulated and real datasets.
Keywords: Empirical Bayes, Fairness, Machine Learning, Variational Inference, Logistic Regression.
Scheduled
GT11.BAYES2 Invited Session
November 7, 2023 6:40 PM
CC1: Audience