Fairness in supervised classification: a comparison between a naïve, a GAN-based, and some other methods

Details

Supervisors

Saerens, Marco

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Concerns about biased results in machine learning models, especially in supervised classification, have led to extensive research on fairness strategies. This thesis presents a comprehensive comparison of some different fairness methods, including Pre-processing, In-processing and Post-processing approaches, as well as a naïve strategy and a last one based on a Generative Adversarial Network (GAN). The effectiveness of these methods in reducing discrimination without significantly compromising accuracy is evaluated on five datasets. The experiments show that the Pre-processing, In-processing and Post-processing strategies show promising results in reducing discrimination while maintaining reasonable accuracy. The naïve strategies show interesting but unreliable results in the reduction of the discrimination. The performance of the GAN-based strategy, although inconsistent, offers opportunities for improvement.