Fairness in Machine Learning

Fairness in machine learning is an increasingly critical topic as AI-driven systems become more embedded in decision-making processes that impact individuals and society. Research on fairness remains fragmented across disciplines, including social sciences, law, and machine learning, and many existing books either focus on theoretical aspects or provide applied ML techniques without addressing fairness-related pitfalls. This book fills that gap, offering a structured and balanced examination of fairness in ML that covers conceptual definitions, fairness-aware algorithms, trade-offs, open challenges, and the regulatory landscape, with clarity on how fairness is measured and the tensions between different fairness criteria. The book is structured to first introduce fairness as a concept, outlining its various definitions and disciplinary perspectives, including philosophical, legal, economic, and social viewpoints, as well as causal and counterfactual notions of fairness. It then explores approaches to mitigating bias in ML, from pre-processing techniques that adjust datasets to in-processing and post-processing methods. Special attention is given to the fairness challenges that arise in generative models, an area that has gained prominence with the rise of large language models and text-to-image systems. The book also discusses the European regulatory landscape, including the AI Act and non-discrimination law, and how these legal frameworks relate to technical notions of fairness. Readers will find particular interest in the discussion of fairness trade-offs and theoretical results, such as the price of fairness and incompatibility results between fairness criteria, which highlight the complex decisions practitioners must make when striving for fairness while maintaining model accuracy and interpretability. The book also emphasizes open challenges, such as the lack of diverse annotated datasets, the complexities of intersectionality, and fairness under distribution shift. By reading this book, researchers, practitioners, and policymakers will gain a deeper understanding of fairness in ML, learning not just about existing solutions but also about the nuances and limitations of current approaches, and will be equipped with both theoretical knowledge and practical tools to assess fairness considerations in their ML models. A background in machine learning and statistics is recommended for readers to fully grasp the technical discussions.