Fundamental Component Modeling within Recommender Systems

The proliferation of online platforms has significantly increased the scope and complexity of recommender systems. As users engage with diverse content, products, and social groups, it is increasingly important to develop models that can capture the nuanced relationships between users, items, and their interactions. This thesis proposes a comprehensive framework for modeling the core components of recommendation: users, items, and interactions. In user modeling, two published papers explore how users’ participation in groups and social interactions reveal latent interests. I propose two novel methods: one that disentangles these interests to improve recommendation personalization, and another that captures users’ dual roles as initiators and participants in social commerce contexts. In item modeling, a published paper focuses on enhancing recommendation diversity using knowledge graphs (KGs), introducing new techniques to balance accuracy and diversity. One paper focuses on interaction modeling, aiming to optimize Click-Through Rate (CTR) prediction by capturing the dynamic interactions between users and items. All experiments are conducted on public datasets (e.g., Steam, Amazon, MovieLens) and an industrial dataset from Meta, demonstrating the effectiveness of the proposed models across different domains.