Semantic-enhanced Heterogeneous Graph Contrastive Learning for Recommendation

Abstract Heterogeneous Graph Neural Network (HGNN) has become increasingly indispensable in recommender systems due to its capability to integrate multi-source heterogeneous information, thereby modeling real-world recommendation scenarios effectively. Concurrently, contrastive learning has achieved significant success in graph learning. However, most existing contrastive learning methods utilize random perturbations to enhance user/item embedding representations. This unreliable data augmentation strategy often leads to the loss of crucial graph structure and introduces misleading supervision signals, thereby hindering the effective capture of rich semantic information within heterogeneous graphs. To address these issues, we propose a novel model named Semantic-enhanced Heterogeneous Graph Contrastive Learning for recommendation (SeHGCL). SeHGCL designs a generative reconstructed data augmentation method with Masked Auto- Encoder (MAE) to capture the intrinsic semantic connections within the graph structure, thus obtaining the semantic-enhanced embeddings and addressing the unreliability of traditional data augmentation methods. Besides, the supervision signals derived from this generative reconstructed data augmentation are then utilized to aid the learning of the primary task’s representation through an enhanced dual-view contrastive learning method. Experiments conducted on three real-world datasets demonstrate that SeHGCL can effectively enhance the semantic representation with respectable performance over state-of-the-art baselines. Codes are available at https://github.com/jerrica0409/SeHGCL