协同信号增强的大模型用户画像生成与推荐OA

The quality of the user profile directly affects the performance of the recommender system.User profile in a traditional recommender system can be derived through modeling the collaborative information between users and items,but is unable to fully utilize the text description information of users and items.The textual information pro-cessing and commonsense reasoning capabilities of LLMs,combined with their world knowledge,provide new oppor-tunities for user profiling.The combination of a recommender system and LLM can give full play to the advantages of each other,and improve each other's performance mutually.This paper proposes a method to introduce two collaborat-ive signals,named potential interest and collaborative scale,from a recommender system into LLM to further enhance the user profile generation of LLM.The user profile is generated through multiple times of interaction with LLM,and is further transformed into an embedding,fusing with the user representation in the recommender system through contrast-ive learning to improve recommendation performance.Experimental results on two datasets and multiple recommender models show that the proposed method can significantly improve the performance of the recommender model.The pro-posed method bridges the gap between LLM and recommender system,and sheds light on further similar research work.