SetLLM: Set Large Language Model for Cold-Start Item Recommendation

Cold-start item recommendation remains a longstanding challenge, as cold items usually lack sufficient interactions to train the behavioral embeddings. Recently, large language model (LLM)-based item cold-start models have tried to address this issue by generating synthetic user-item interactions to train the behavioral embeddings for cold items. Current cold-start LLM paradigms, such as PairLLM, generally first select a small set of users and then filter possible users for the cold items as cold interactions. However, PairLLM suffers from linearly increasing computational complexity as the user candidate set grows, resulting in substantial computational overhead and limited performance. In this paper, we propose a novel SetLLM paradigm, which directly generates a potential interaction subset rather than inferring interactions individually from a user candidate set. Specifically, SetLLM first constructs a subset generation structure that utilizes the content of cold items to filter the user candidate set according to user behaviors and then incorporates tailored set generation fine-tuning mechanisms. Moreover, we propose a Direct Cold-start Preference Optimization (DCPO) module to evaluate and refine the quality of the generated cold-start interaction sets, further enhancing the performance of SetLLM. Together, these components establish an effective paradigm for LLM-based cold-start recommendation. Experiments on two real-world datasets demonstrate that SetLLM achieves state-of-the-art performance on item cold-start recommendation while improving computational efficiency by over 30 times compared with the PairLLM approach.