The Open University
Enhancing Fairness in Machine Learning: Identifying and Mitigating Bias with a Focus on Gender Bias in Finance
Abstract
dc:description.abstractThe integration of Artificial Intelligence (AI), particularly Machine Learning (ML), has revolutionised processes across various domains by automating decision-making tasks such as detecting illnesses in healthcare or assessing credit risk in finance. However, ML models are trained on data that may reflect historical and societal biases, which these models can learn and propagate. This perpetuation of bias poses significant risks, particularly in high-risk applications that impact individuals' well-being, such as the ones used in healthcare and financial services. Without proper intervention, these systems may reproduce or even exacerbate existing inequalities and disproportionately harm marginalised groups. This thesis delves into three core challenges related to bias in ML models: identifying information that biases ML models and using it to mitigate such bias, automatically mitigating bias in the absence of sensitive information (e.g. gender), and addressing multiple fairness definitions simultaneously when mitigating ML bias. First, the thesis introduces the Covariance Proxy Identification approach, a novel approach to uncovering attributes within datasets that act as proxies for sensitive information (e.g. maternity leave as a proxy for gender). Using statistical tests of covariance, the approach identifies indirect representations of sensitive attributes that can propagate bias to ML models. By removing these proxies, the approach demonstrates a significant reduction in bias in ML models while maintaining high levels of predictive performance. Evaluations on real-world financial datasets validate the utility of this approach, offering a pathway to address bias rooted in the data itself. Second, addressing the frequent unavailability of sensitive information due to privacy, ethical, or legal constraints, the thesis proposes a knowledge transfer approach. This approach automates the identification of bias-inducing proxies in datasets where sensitive attributes are accessible and transfer this knowledge to analogous scenarios where such attributes are unavailable. Experiments using U.S. Census data highlight the efficacy of this approach, achieving fairness improvements in challenging scenarios where demographic data is missing. Third, the thesis explores the interplay between competing fairness definitions, a significant challenge in creating fair ML systems. By systematically evaluating 26 distinct bias mitigation approaches and their combinations across six fairness metrics, the research uncovers effective strategies for balancing conflicting fairness definitions. The findings demonstrate that combining bias mitigation approaches can achieve greater alignment with multiple fairness definitions. Additionally, the work provides actionable guidance and recommendations for bias mitigation combinations while introducing an open-source benchmark to facilitate fairness evaluation across multiple constraints. While the proposed approaches and experiments are generalisable across domains, this thesis specifically focuses on addressing ML gender bias in the financial domain due to the historical bias in financial data, particularly against women. Through these contributions, the thesis advances the field of responsible AI, offering actionable solutions and recommendations to identify and mitigate bias in ML models, ultimately fostering more fair AI systems.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pavón Pérez, Ángel
Rights
- Language dc:language
- en