DPRS: Distributional Perspective Modeling for LLM-Enhanced Recommendation Systems

Existing recommender systems mainly rely on ID embedding to model user behavior sequences. However, this approach ignores the large amount of semantic information embedded in the textual descriptions of items, resulting in low performance. The ability to utilize large-scale language models to understand and reason about textual content offers a promising avenue for recommender system development. To this end, we propose a recommender system model, DPRS, which models textual representations of large-scale language models from a data distribution perspective and incorporates them into traditional recommendation models. We utilize pre-trained LLMs as item encoders to introduce textual representations to the recommender system. To address the heterogeneity between the LLM representations and the original recommender system, we propose to model user behavior sequences from a data distribution perspective by converting them into similarity score sequences. By conducting experiments on two public datasets, we verify the superiority of the proposed model.