color:#002060">Machine learning is increasingly deployed in domains where predictions directly affect individuals, including hiring, credit, healthcare, education, and public administration. Recently, the rapid adoption of large language models has further expanded the scope and societal impact of machine learning. In this context, concerns about the fairness of machine learning systems have become a major scientific and societal issue, as models trained on historical data may reproduce or amplify existing social inequalities. Yet fairness remains difficult to define, assess, and enforce. It is not a single statistical property of model outputs, but a normatively loaded and context-dependent concept whose meaning depends on the type of disparity one aims to prevent. Measuring such a concept through a single scalar indicator is therefore inherently risky. The resulting score may be misaligned with the underlying fairness objective while concealing this mismatch behind an apparently objective numerical value. Current fairness methods have been developed within a fragmented landscape of fairness definitions and metrics, with most approaches tied to a single fairness definition and designed for simplified settings such as binary sensitive attributes and binary classification. As a result, methods remain limited in their ability to accommodate the diversity of practical applications, particularly intersectional settings involving multiple sensitive attributes and tasks with multiclass outputs. More broadly, the relationship between normative fairness objectives and the mathematical criteria used to operationalize them often remains insufficiently explicit, making fairness assessment both technically challenging and conceptually fragile. Addressing these limitations is essential for developing robust, interpretable, and operational mathematical tools for fairness-aware machine learning.
color:#002060">This thesis seeks to address the limitations of current methods by developing a normatively grounded mathematical framework for assessing fairness and mitigating biases in machine learning. The first contribution, MIFair, is a unified information-theoretic framework for fairness assessment and in-processing mitigation. It introduces the concept of benefit as a common formal intermediary underlying several established group fairness notions and uses mutual information to measure the statistical dependence between sensitive attributes and this benefit variable. This yields a flexible metric template that recovers multiple fairness definitions within a single formalism, facilitating systematic metric selection aligned with intended fairness while naturally extending to intersectional and multiclass settings. The same formulation is then integrated into a regularized empirical risk minimization framework, providing a corresponding in-processing mitigation method. The second contribution focuses on the construction of scalar fairness scores in multi-group settings, including intersectionality. It identifies and formalizes meta-metrics as a distinct and overlooked design layer in fairness evaluation, corresponding to how subgroup-level disparities are aggregated into a single score. It proposes a taxonomy of these meta-metrics, clarifies the normative implications of their design choices, and shows that different aggregation strategies may lead to substantially different fairness judgments. This analysis is further extended through practical guidelines to guide practitioners in selecting an appropriate meta-metric, thereby supporting the broader objective of maintaining alignment between what is measured and the fairness concept it is intended to represent.
color:#002060">Together, these contributions provide a principled mathematical framework for assessing, mitigating, and interpreting fairness in machine learning across a broad range of practical settings. The proposed methods are supported by theoretical analyses and experimental evaluations on publicly available datasets.
