Towards Trustworthy Personalized Machine Learning Systems under User-System Interactions

Abstract

The revolutionary wave of Industry 5.0 heralds a new era of human-centered industrialization that moves beyond automation-centered production. Cutting-edge technologies, including artificial intelligence/machine learning (AI/ML) and big data, are increasingly expected to support human welfare and societal needs. In response to this emerging trend, the interests and experiences of real users have become central to the design of machine learning systems across various applications, such as healthcare, transportation, supply chains, etc. To build such ML systems, there are three critical factors to consider: 1) User Heterogeneity: Users differ in their needs, preferences, behaviors, and contexts. Therefore, ML systems should explicitly account for such heterogeneity and personalize services for individual users so as to improve service quality and user satisfaction. 2) System Trustworthiness: The real-world effectiveness of these personalized ML systems relies on sustained user engagement and feedback, which are grounded in user trust. These systems must therefore be trustworthy, protecting users’ interests and mitigating potential risks arising from personalized services through characteristics such as robustness and fairness. 3) User-System Interactions: Instead of one-way delivery, ML systems should interact with users to collect their feedback and enhance their engagement under their behavioral uncertainty or environmental complexity. These factors highlight the system-level challenges central to building trustworthy personalized ML systems that engage heterogeneous users. To address these challenges, this dissertation contributes AI/ML methodologies that integrate these three aspects. Chapter 2 focuses on the representation of user heterogeneity at scale and presents CrowdLLM, a framework that combines large language models (LLMs) and generative models to create high-fidelity digital populations for more cost-effective decision-making. Chapter 3 and Chapter 4 center on two key dimensions of trustworthiness: fairness and robustness. Chapter 3 develops fair collaborative learning (FairCL), a framework that can incorporate a variety of fairness concepts to improve fairness amid personalization, along with a self-adaptive algorithm for effective implementation. Chapter 4 further studies user-system interactions in reward systems and presents a robust learning method for modeling user choice behaviors under such interactions. Taken together, this dissertation takes a step toward advancing trustworthiness and personalization in ML systems under user-system interactions. Its contributions lay a foundation for broader efforts to develop human-centric AI-enabled systems that can adapt to user disparities and support complex decision-making with real-world settings in the future.