Maggie Wu's Preliminary Exam

Abstract: As machine learning systems are increasingly deployed in high-stakes decision-making domains such as finance, healthcare and recommendation, ensuring fairness across demographic groups has become a critical societal and technical concern. Efforts to mitigate bias often come at the cost of predictive performance, posing a persistent challenge: how can we harmonize fairness and utility in machine learning systems? Despite growing interest in algorithmic fairness, practical deployment remains challenging due to real-world constraints such as incomplete labels, limited access to sensitive attributes, and severe group imbalance. Moreover, fairness is often treated as a static objective, with limited understanding of how it evolves during training. In response, this thesis proposes a unified research agenda to harmonize fairness and utility in machine learning, spanning both data-centric and model-centric perspectives. From the data perspective, I develop learning algorithms that identify and correct representational imbalances in the training data, enabling more equitable predictions across demographic groups. These approaches, including reweighting and recalibration, improve group-level fairness without extensive labeling or sensitive group annotation. From the model perspective, I propose fairness-aware learning objectives that harmonize fairness with utility during training, such as incorporating fairness regularizers and debiasing optimization mechanisms in neural models. Across both directions, I evaluate the proposed methods on a range of benchmark tasks and real-world applications, demonstrating improvements in group fairness metrics while maintaining strong overall performance. Together, these contributions aim to advance the development of fair, reliable, and practically deployable machine learning systems, and deepen the theoretical understanding of fairness in learning systems.

Questions? Contact Maggie Wu