CKEN: Collaborative Knowledge-Aware Enhanced Network for Recommender Systems

Abstract

AbstractKnowledge graph are widely used as a kind of side information to alleviate the data sparsity in collaborative filtering. However, a majority of existing recommendation methods incorporating knowledge graph emphasize how to encode knowledge associations in the knowledge graph, while ignoring the key collaboration information implied in the user-item history interactions and the shared information between the item and the knowledge graph entities, resulting in insufficient embedding representation of users and items. To that end, we propose a collaborative knowledge-aware enhanced network (CKEN). Specifically, CKEN uses a collaborative propagation approach to explicitly encode user-item collaborative information and then allows it to be propagated in the knowledge graph, and applies an attention mechanism to obtain contributions from neighbors with different knowledge during knowledge propagation. In addition, considering the association relationship between items and knowledge graph entities, this paper designs a feature interaction layer to facilitate feature sharing between items in the recommender systems and entities in the knowledge graph in order to further enrich the latent vector representation of items. Numerous experiments on several real-world scenario datasets have shown that CKEN’s recommendations are significantly better than several other best-of-class baselines. KeywordsRecommender systemsCollaborative propagationFeature interactionKnowledge graph