Synergizing Large Language Model and Knowledge Graph for Cold-Start Recommendation

The cold-start item problem in recommendation systems has long been challenging and widely studied because there is no interaction data for cold-start items. This makes it difficult to obtain effective representations of these items and leads to a decline in the recommendation performance. The state-of-the-art methods in cold-start recommendation mainly leverage LLMs to extract relevant semantic information from the content of cold items to enrich their representation. However, LLM-based methods often generate incorrect information due to the lack of sufficient factual knowledge and groundings. Instead, KG contains factual knowledge about users and items and can provide LLMs with more comprehensive groundings. Therefore, we propose a novel framework, namely SCSRec, which synergizes both LLM and KG to generate high-quality interaction simulation data for cold items and further enhances the performance of cold-start recommendations. It mainly generates high-quality cold item representations with the synergy of LLM and KG for interaction simulation and uses the interactions to train the recommendation model. Specifically, first, to obtain high-quality representations, we synergize the LLM and KG using a dual-enhanced method. We use the LLM to enrich the descriptions of entities and complete the structural information of cold items to obtain a KG with rich information. Then we extract entities’ sub-graphs as factual knowledge and groundings for LLM representation generation. Next, we introduce an Adaptive Expert Synergy Module to deeply integrate and synergize the LLM-based representations and the KG-based representations of users and items. Based on the integrated high-quality embedding, we use the similarity between items and users to generate simulated interactions for cold items. Finally, we combine the simulation data with the interaction data of warm items from the dataset to jointly train the recommender and enhance the performance of cold-start recommendations. Experimental results on three public datasets show that our model outperforms existing state-of-the-art methods in the scenarios of warm, cold, and overall recommendations.