A Knowledge Graph Recommendation Approach Incorporating Contrastive and Relationship Learning

Abstract

Knowledge graphs, encompassing abundant structured and semantic information, exhibit remarkable potential within personalized recommendation fields. However, most current recommendation approaches reliant on knowledge graphs are susceptible to the disturbances of noise and data sparsity present in low-quality knowledge graphs, thereby undermining the efficacy of these recommendations. We propose a recommendation model grounded on contrastive learning and item relation learning to address these challenges. Specifically, to alleviate the issue of noise interference in knowledge graph feature learning and enhance model robustness, we introduce a hierarchical contrastive learning method. This approach leverages contrastive learning within knowledge graphs to extract unique features from the same propagation layer, thereby mitigating the impact of noise. In the face of the data sparsity endemic to low-quality knowledge graphs, we propose an item relation learning strategy that uses data augmentation to generate unknown items and their interrelations, hence enriching the content of the knowledge graph. We evaluate our model utilizing two public datasets, and the results, in comparison to other state-of-the-art methodologies, indicate a superior recommendation performance under conditions of a long-tail issue and data scarcit.