As machine learning systems are increasingly deployed in high-stakes decision-making contexts, concerns about algorithmic bias and unfairness have come to the forefront. Achieving algorithmic fairness requires coordinated advancements in how we define...
As machine learning systems are increasingly deployed in high-stakes decision-making contexts, concerns about algorithmic bias and unfairness have come to the forefront. Achieving algorithmic fairness requires coordinated advancements in how we define fairness (metrics), enforce it (learning algorithms), and assess it (evaluation datasets). However, existing approaches often struggle to generalize across tasks, scale to realistic settings, or address fairness in underexplored domains such as generation tasks.
This dissertation addresses these limitations by proposing a comprehensive investigation into novel fairness metrics, fairness-aware algorithms, and benchmark datasets. First, we propose a fairness-aware distillation framework that aligns group-conditional representations between teacher and student models using a regularization term based on Maximum Mean Discrepancy (MMD), improving both accuracy and fairness in visual recognition tasks. Second, we introduce a reweighting-based optimization method grounded in classwise Group Distributionally Robust Optimization (DRO), which reduces group-level disparities for each class. Third, we present a more practical fairness learning framework that utilizes confidence-based pseudo group labels to enable fairness-aware training when only partial demographic information is available in the training dataset. Fourth, we construct novel benchmark datasets of counterfactual images and theoretically analyze the relationship between counterfactual fairness and group fairness in image classification, revealing their divergence and offering a new method to bridge them. Finally, we propose a new metric—Multi-Group Proportional Representation (MPR)—to evaluate representational fairness in generative models, providing a scalable and flexible measure across intersectional groups and prompt-specific contexts.
Through this body of work, we offer rigorous and useful tools to advance the algorithmic fairness of machine learning models across diverse modalities and more practical real-world applications. Our findings underscore the need to move beyond simplistic fairness paradigms and toward principled, scalable, and context-sensitive solutions that better reflect the complexity of real-world AI deployments.