More about HKUST
Fairness-alignment in Data-centric Machine Learning: From Regularization to Equilibrium
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Fairness-alignment in Data-centric Machine Learning: From Regularization to Equilibrium" By Miss Yue CUI Abstract: Ensuring fairness in machine learning is crucial for ethical decision-making and fostering societal trust. However, achieving fairness across diverse, real-world datasets presents significant challenges. Traditional controlling-based methods, such as fairness regularization, have been widely explored but often lack generalizability and struggle to accommodate multiple protected attributes or adapt to complex, heterogeneous data distributions. This thesis first addresses these limitations by proposing a universal fair representation learning framework that captures fairness w.r.t. an exponential number of constraints. Building on this foundation, we further explore fairness in dynamic and personalized settings. To enhance adaptability, we introduce a personalization-based approach utilizing a pre-train and fine-tune framework. This method ensures uniform performance across diverse target groups while addressing fairness misalignment in individualized scenarios. However, personalized fairness introduces new challenges, including susceptibility to strategic manipulation and conflicts among stakeholders with differing interests. To resolve these issues, we develop an incentive-based fairness alignment paradigm that integrates equilibrium-driven strategies. By incorporating momentary rewards into fairness mechanisms, this approach balances fairness objectives with stakeholder incentives, fostering more sustainable and robust fairness solutions in real-world applications. The findings of this thesis contribute to a more comprehensive understanding of fairness in machine learning, advancing both theoretical insights and practical applications in ethical AI development. Date: Wednesday, 26 March 2025 Time: 2:00pm - 4:00pm Venue: Room 5504 Lifts 25/26 Chairman: Dr. Ding PAN (PHYS) Committee Members: Prof. Xiaofang ZHOU (Supervisor) Prof. Cunsheng DING Prof. Qiong LUO Dr. Can YANG (MATH) Prof. Jeffery Xu YU (CUHK)