Knowledge Graph-based Recommendation via Generative Denoising and Multi-Level Contrastive Learning

Integrating knowledge graphs into recommendation systems poses significant challenges, particularly in handling noise and distribution mismatch between the knowledge graph and the user–item interaction graph. To address these challenges, we propose a knowledge graph recommendation framework that combines generative denoising with multi-level contrastive learning. This integration of complementary self-supervised objectives alleviates the lack of supervision and enhances the model's ability to learn from sparse user-item interactions. Specifically, to handle the high heterogeneity of knowledge graphs, our method employs a heterogeneous graph Transformer for representation learning. To suppress noise in the graph structure, we introduce a credible knowledge masking and reconstruction mechanism. Furthermore, for fine-grained user preference modeling, we leverage decoupling strategies and multi-preference contrastive learning to capture multi-dimensional interests, enhancing the interpretability of user behavior. Extensive experiments demonstrate that our method surpasses the state-of-the-art baselines