Machine learning methods are increasingly being utilised as decision-making tools across various domains. However, concerns related to fairness have been raised regarding their use in decision-making processes. In response, many researchers have developed approaches to address these concerns, including de-biasing algorithms and formal notions of fairness that define the conditions for algorithmic fairness. This thesis explores one such notion, known as Max-Min Fairness. While numerous algorithms have been proposed to support Max-Min Fairness, this thesis demonstrates that classic boosting algorithms, particularly the AdaBoost algorithm, are capable of supporting Max-Min Fairness despite not being originally designed for this purpose. In fact, boosting algorithms often outperform previous approaches that were specifically designed to support Max-Min Fairness. We analyse the sample weights generated by the AdaBoost algorithm to explain how it facilitates its support of Max-Min Fairness and explore methods for modifying or enhancing these weights to further improve the extent to which Max-Min Fairness is supported. Additionally, we investigate common challenges within the field and show how they can be fully or partially mitigated when boosting algorithms are applied to fairness-related tasks in machine learning.
Boosting Algorithms for Max-Min Fairness in Machine Learning
Ziubroniewicz, D. (Author). 8 Oct 2025
Student thesis: Doctoral Thesis › PhD