ML-KGCL: Multi-level Knowledge Graph Contrastive Learning for Recommendation

The knowledge graph-based (KG-based) recommender systems have achieved excellent results in the recommendation domain. However, the long-tail issue hinders the model from mining the real interests of users. Existing research has shown that Contrastive Learning (CL) can alleviate the long-tail issue, but the existing graph contrastive learning methods are not completely compatible with KG-based recommendation. To fill this gap, we propose a Multi-Level Knowledge Graph Contrastive Learning framework (ML-KGCL) to introduce CL into the KG-based recommendation. ML-KGCL makes the CL task more compatible with the recommendation task while mitigating the long-tail issue by performing fine-grained node representation learning. Firstly, we generate positive samples via graph augmentation strategy. Then, we divide the KG-based recommendation into three levels: user-level, entity-level and user-item-level, and perform fine-grained multi-level CL to optimize the node representations. Next, we obtain the final node representations through the signal integration strategy. Finally, the model is trained by the joint learning paradigm. The experimental results on three public datasets are better than the baseline models.