New Community Cold-Start Recommendation: A Novel Large Language Model-based Method

Shangkun Che, Minjia Mao, Hongyan Liu

Abstract: The New Community Cold-Start problem refers to the serious challenge of building recommender system faced by a newly established community, where there is a lack of users and user engagement. To tackle this problem, we propose a novel recommendation method, leveraging the inference capability of LLM. We choose In-Context-Learning as our prompting strategy and design a coarse-to-fine framework to choose demonstration examples to create effective ICL prompts efficiently. We leverage the bias exhibited by LLM in perceiving prompts in the coarse stage and a reinforcement learning model to learn an example selection policy in the fine stage. Empirical studies demonstrate our method can greatly improve the performance of recommendation for the NCCS problem.