Knowledge Graph Self-Supervised Learning for Recommendation

Xinye Guan, Guixin Chu, Bangzuo Zhang, Lingxuan Li

Abstract: The integration of knowledge graphs (KG) into recommender systems has been proven to effectively alleviate data sparsity. A recent technical trend in knowledge-aware studies is to develop multi-task learning models founded on contrastive learning (CL). However, the current CL-based methods often focus on static knowledge associations, overlooking dynamic user interests, and thus limiting the scope of performance improvement. To address this limitation, this paper proposes a novel Knowledge Graph Self-Supervised Learning (KGSSL) method. We first propose a novel method for modeling structural importance for knowledge triplets which can adaptively identify informative knowledge associations by learned structural score for KG triplets. Furthermore, to capture users' dynamic interests, we propose a novel fact-guided variational graph generation task. By injecting structural scores into the variational graph autoencoder, KGSSL is trained to capture users' latent behavioral patterns. To effectively leverage both users' dynamic interests and factual knowledge associations, we also utilize contrastive learning to align the signals between the knowledge graph and the user-item graph. Ideally, this can help boost GCL using more diverse contrastive views while leveraging static reliable knowledge to guide the graph distribution learning. Extensive comparison and ablation experiments on five realworld datasets demonstrate that KGSSL outperforms state-of-the-art methods

External IDs:dblp:conf/ijcnn/GuanCZL25