A Model Hybrid Recommendation Approach Based on Knowledge Graph Convolution Networks

In previous works, researchers proposed the knowledge graph (KG) and knowledge graph convolution networks (KGCN) for mitigating sparsity and cold start complications present in collaborative filtering recommender systems. It is noted that KGCN is mainly used to capture the correlations between items; they generally ignore the correlations between users, thus resulting in insufficient accuracy of recommendation outcomes. In this study, we developed a model hybrid recommendation approach based on knowledge graph convolution network collaborative filtering (KGCN-CF). This approach adds user-based collaborative filtering based on KGCN to capture the correlations between users, which counterbalances the defects of KGCN, while addressing the sparsity and cold start complications of collaborative filtering to a certain extent. The proposed approach was applied to three datasets provided by Dianping-Food, Book-Crossing, and Last.FM. The experimental results demonstrated that this approach can effectively improve the area under the curve (AUC) and F1-measure of the recommendation.