An empirical study of bias and discrimination induced by different machine learning models

Details

Supervisors

Saerens, Marco

Faculty

Ecole polytechnique de Louvain

Degree label

Master [60] en sciences informatiques

Abstract

As machine learning systems become increasingly embedded in decision-making processes—particularly in sensitive domains such as finance, education, and criminal justice—questions of fairness and bias have become central. This thesis explores how different machine learning models can influence existing biases in data, either by amplifying, mitigating, or maintaining them. Rather than focusing solely on techniques for correcting bias, the study examines how model design, complexity, and training choices affect fairness outcomes across diverse datasets and demographic groups. The analysis combines classical performance metrics with fairness indicators to assess the behavior of six widely used classification algorithms. It also considers whether factors such as hyperparameter tuning or changes in bias for one protected group systematically affect others. The findings highlight that fairness is a multidimensional challenge that cannot be resolved by algorithm choice alone. Instead, it requires a holistic approach, combining ethical considerations, data integrity, and thoughtful model evaluation to develop responsible and equitable AI systems.