Towards the Next Generation of Recommendation Systems: Exploration, Trustworthiness, and Context-Aware Generation

Abstract

Recommender systems have become indispensable tools for navigating information overload, achieving remarkable success in personalizing content across domains like e-commerce, healthcare, and entertainment. However, this success has also exposed critical gaps: modern systems often promote content homogenization over diversity, operate as opaque "black boxes" that are difficult to interpret, and risk perpetuating societal biases through unfair predictions. These challenges underscore the urgent need for a new generation of recommender systems that are not only accurate but also robust, explainable, and contextually adaptive. This thesis advances this vision by grounding its contributions in three foundational pillars: Exploration, Trustworthiness, and Context-Aware Generation. First, under the exploration pillar, we develop complementary recommendation strategies using logical reasoning. This approach moves beyond traditional similarity-based methods to construct network architectures that balance immediate user satisfaction with the discovery of diverse and novel items. Second, we focus on trustworthiness by addressing both calibration and fairness under uncertainty. We propose a novel metric to align model confidence with accuracy at the individual level and provide a theoretical framework for ensuring fairness in uncertain prediction scenarios. Third, to promote trust and transparency through context-aware generation. We introduce a neural architecture search framework for explainability. This method identifies interpretable user-item interaction patterns via cross-attention mechanisms, making recommendations clearer and more intuitive. Additionally, we propose a framework that utilizes the exponential volume expansion of hyperbolic space to index items into structured semantic IDs, enabling generative models to better capture hierarchical relationships and improve performance for sparse, niche content. We further present an adaptive trend-aware recommendation framework that transforms heterogeneous external signals into actionable recommendation surfaces through multi-source trend detection, topic linking, hybrid content retrieval, and controllable presentation interfaces. By incorporating source-grounded explanations and editorial control, this framework enables recommender systems to respond to emerging real-world events in a timely, interpretable, and operationally practical manner. This thesis lays the groundwork for building adaptive, responsible, and next-generation recommender systems that serve diverse user needs through principled exploration, robust design, and contextually intelligent generation.