More about HKUST
Fairness-alignment in Data-centric Machine Learning: From Regularization to Equilibrium
PhD Thesis Proposal Defense Title: "Fairness-alignment in Data-centric Machine Learning: From Regularization to Equilibrium" by Miss Yue CUI Abstract: Fairness in machine learning is essential for ethical decision-making and societal trust, yet achieving it in diverse, real-world datasets remains challenging. Traditional controlling-based methods, such as fairness regularization, have been extensively studied for mitigating bias. However, these methods often lack generalizability and struggle to incorporate a large number of protected attributes or adapt to complex, heterogeneous data distributions. To address these limitations, we propose a universal fair representation learning approach that captures fairness constraints in a more generalized and scalable manner. While regularization-based methods provide foundational solutions, they fail to address fairness alignment in dynamic or personalized settings. To this end, we introduce personalization-based approaches, a pre-train and fine-tune framework, which offer tailored solutions to ensure performance uniformity across diverse target objects. However, these approaches come with new challenges, including vulnerability to strategic manipulation and conflicting fairness priorities among stakeholders. To mitigate these challenges, we explore a novel incentive-based fairness alignment paradigm that emphasizes equilibrium-driven strategies. By integrating momentary rewards into fairness mechanisms, this paradigm balances fairness and stakeholder incentives, paving the way for more robust and sustainable solutions in real-world applications. Date: Wednesday, 5 February 2025 Time: 2:00pm - 4:00pm Venue: Room 4472 Lifts 25/26 Committee Members: Prof. Xiaofang Zhou (Supervisor) Prof. Qiong Luo (Chairperson) Prof. Cunsheng Ding Dr. Lei Li