Exploring indirect entity relations for knowledge graph enhanced recommender system

Knowledge graph (KG)-based recommendation models generally explore auxiliary information to alleviate the sparsity and cold-start problems in recommender systems. Previous approaches enhance representations of users and items by exploring the influence of multi-hop neighbors. However, existing works fail to consider the indirect feedback for improving user representation and the diversity of the multi-hop neighbors for enriching item representation. To this end, we present a novel recommender system, called Entity Relation Similarity and Indirect Feedback-based Knowledge graph enhanced Recommendation (ERSIF-KR) to enhance representation learning in KG-based recommender systems. In addition, our model exploits indirect feedback of items that are not directly interacted with users to alleviate the exposure bias while enhancing user similarity computation when learning user representation. Moreover, our method directly incorporates representation of multi-hop neighbors into the target item embedding with weights determined by the correlations between high-order and low-order relations, which can significantly boost the item representation learning. Extensive experiments on three real-world datasets demonstrate that our model achieves remarkable gains in terms of recommendation performance and model convergence time, and effectively alleviates the sparsity and cold start problems.

Automated summary

Knowledge graph-based recommendation models use auxiliary information to address sparsity and cold-start problems. However, existing approaches neglect indirect feedback and diversity of multi-hop neighbors. We propose a novel recommender system that enhances representation learning by considering these factors.