Knowledge-enhanced recommendation using item embedding and path attention

摘要

Recommender systems have attracted widespread attention in various online applications. To effectively recommend the needed items of users, knowledge graphs have been introduced to provide rich and complementary information to infer user preferences in recommender systems. Existing efforts have explored user preferences through specific paths and item embedding in knowledge graphs. However, user preferences have hardly been fully captured because users and items are always separately modeled. To address this problem, we propose a model to represent items from a user’s perspective that provides effective supplementary information. User preferences encoded in historically clicked items are propagated along links in the knowledge graph. We propose a gated attention unit to capture user preferences from specific types of paths. Based on the captured preference information through the knowledge graph and supplementary item information, we generate effective reasoning paths to infer the underlying rationale of user–item interactions using the sequential model. Through extensive experiments on real-world datasets, we demonstrate that the proposed model achieves significant improvements over the state-of-the-art solutions.