Fairness in machine learning : focus on post-processing methods

Details

Supervisors

Saerens, Marco

Faculty

Louvain School of Management

Degree label

Master [60] en sciences de gestion

Abstract

The popularity of machine learning derives from its application to decision-making in all fields. Classification is fundamental, as it allows data to be categorized on the basis of characteristics, with supervised classification learning from labeled data to predict outcomes. The question of fairness is essential to avoid biased results and promote equal treatment for all in many areas such as clinical testing or criminal justice. There are a number of different fairness measures, each more appropriate in certain situations. Promoting fairness in machine learning is crucial for responsible and fair decision-making. This thesis develops a new fair post-processing method, Optimal Swapping, and its greedy version, for binary score predictions in supervised classification. It aims to mitigate unfairness while maintaining overall accuracy. Comparison analyses with other post-processing techniques, using statistical tests such as Friedman, Nemenyi, and Wilcoxon signed-rank, evaluate the method's efficiency in promoting fairness. By contributing to robust and effective fairness techniques, this work advances fairness machine learning applications.