Leveraging hierarchy-aware diffusion model and knowledge-enhanced contrastive learning for recommendation

Kaibei Li, Yihao Zhang, Qinyang He, Xiaokang Li

Abstract: The knowledge graph (KG) provides a reasonable explanation for comprehending the semantic relationships among items in recommendation systems. Recently, there has been interest in harnessing knowledge graph-enhanced contrastive learning to improve data representation in recommendations. However, most methods rely heavily on manually generated views for contrastive learning, which hinders their applicability to downstream recommendation tasks. Additionally, these methods often ignore the issue of knowledge imbalance, potentially disrupting the true intentions of users. To address these challenges, we propose a hierarchy-aware diffusion model and knowledge-enhanced contrastive learning for recommendation. Technically, we first introduce a hierarchy-aware diffusion model in hyperbolic space. This approach combines the masked prediction paradigm with the diffusion model, effectively enhancing the adaptability of generated views to downstream recommendation tasks. Furthermore, we introduce a diffusion cone constraint loss to filter out semantically irrelevant generated data. Additionally, we design a local topology-aware aggregation strategy that effectively extracts both local and global node features. In this way, we achieve the unification of hierarchy-aware diffusion model and knowledge-enhanced contrastive learning, which can effectively address knowledge imbalance and enhance adaptability to downstream recommendation tasks. Our experiments on four datasets show that our model outperforms the state-of-the-art baselines. Our code is available at https://github.com/likaibei/ReSys_HDKCL.

External IDs:dblp:journals/kais/LiZHL26