Enhancing Personalized Recommendations with Transferable User Representation Learning in Limited Data Contexts

Jing Du

Abstract: Confronted with the vast volume of data, Recommender System~(RS) proves to be a valuable tool for information filtering. It comprehensively integrates diverse user-related factors to generate precise user representations, also known as user interests or preferences. This approach facilitates the suggestion of useful items from a large array of available options. Hence, comprehensive investigations are essential to ensure accurate recommendations. Despite the abundance of data, recommender systems challenges arise from disparities in user engagement and highly uneven data distribution across different subjects, known as long-tail distribution. Different engagement results in a situation where a minority of users or items accumulate a large volume of interaction data, leaving the vast majority facing the data sparsity problem. Moreover, the imbalanced distribution often neglects the preferences of newcomers, aggravating cold-start issues in RS. In this thesis, we aim to extract transferable user representations for personalized recommendations in limited data contexts. We specifically list and aim to address the following key challenges: (1) Learning complex patterns with data sparsity in multi-task recommendation; (2) Incorporating knowledge for new users in social recommendation; (3) modeling transferable underlying patterns in sequential recommendation; (4) Achieving non-overlapping user representation matching in cross-domain recommendation; and (5) Ensuring causal representation identifiability in cross-domain recommendation. Our significant contributions can be summarized as follows: First, we investigate task-specific and task-shared representations in multi-task learning and introduce a novel Multi-Task Learning model by explicitly identifying global commonalities and local features of tasks for multiple objectives. Second, we capture user representations by exploring the intrinsic relationships between and within entities, and transfer them to new users using the Socially aware Dual-Contrastive learning framework for Recommendation, named SDCRec. Third, we analyze user preferences from a functional perspective and introduce an interest dynamic modeling framework, named IDNP, to achieve cold-start recommendation based on generative Neural Processes. Fourth, we introduce a cross-domain invariant Preference Matching method, DPMCDR, to capture the continuity in user behaviors within each domain and discover transferable invariant preference across domains. Fifth, we further propose an extension of DPMCDR, named CausalID, to identify relationships between user representation distributions and facilitate knowledge transfer across domains from the causal perspective.

External IDs:dblp:phd/basesearch/Du24