Exploring Entity-level User Preference on the Knowledge Graph for Recommender System

Knowledge graphs (KG) have attracted extensive attention in recommender systems since they contain rich external knowledge. The recent trend in KG-enhanced recommender systems is to employ graph neural networks (GNN) to learn the node representations of the involved graph structures in the recommender system to speculate user preferences. However, existing KG-enhanced recommendation models face two major issues: i) User preferences are mainly built at the relation level and fail to model preferences at the attribute entity level; ii) Implicit feedback lacks accurate user rating information and the data may contain noisy interactions. Such inaccurate preference modeling and imperfect interaction data hinder the capture of users’ actual preferences. To this end, we propose an Entity-level user Preference-aware model on Knowledge Graph (EPKG), which models user preferences at the attribute entity level. Specifically, we introduce the number of connections between attribute entities and user interaction items in the knowledge graph and establish a weight distribution on the number of connections to speculate user preferences for attribute entities. Furthermore, we devise user preference learning to model user preferences to the finer attribute entity level. Afterward, we design a preference-aware aggregation strategy that uses entity-level user preferences to guide the learning of item weights in user interaction history, which in turn alleviates the effects of lack of user rating information and noisy interactions. Experimental results on the three datasets show that EPKG achieves significant improvement compared to the state-of-the-art models. Especially for the Last-FM dataset, EPKG improves NDCG@20 and Recall@20 by 31.5% and 18.4%, respectively.