Behavior Mining for Recommendation

In the era of information explosion, recommender systems are crucial for mitigating user information overload by filtering out irrelevant information and suggesting preferred items or content. The effectiveness of these systems hinges on accurately understanding user prefer- ences through the analysis of vast user behavior records. However, several challenges hinder the optimal performance of recommender systems. Notably, cold-start problems and noise within user behavior records significantly hinder learning user behavior patterns and develop- ing effective recommendation models. Additionally, over-smoothing issues inherent in common recommender system architectures further complicate the creation of intelligent recommender systems. This thesis is articulated through the discussion of four published papers and one paper under review, each addressing these impediments from different perspectives. The first paper tackles cold-start issues through the integration of knowledge graphs, and proposes a learnable sampling strategy to mitigate noise interference from these graphs. The second paper addresses noise in sequential recommendation scenarios via employing Variational Autoencoders (VAEs) to directly capture noise and uncertainty. The third paper overcomes cold-start difficulties by leveraging the knowledge embedded in pretrained large language models, along with a focused capture of user intent dynamics. The fourth paper confronts over-smoothing by incorporating diffusion models into the recommendation processes. Finally, the fifth paper takes a novel approach by utilizing the emergent capabilities of large language models for few-shot inference, offering a broad spectrum approach to the prevailing issues discussed.