Multi-level Disentangled Contrastive Learning on Heterogeneous Graphs

Self-supervised learning methods, including contrastive learning based on Heterogeneous graph neural networks (HGNNs), have achieved great success in learning the representations of heterogeneous information networks (HINs). However, the existing self-supervised methods usually neglect the entanglement of the latent factors behind HINs, which decreases the performance of downstream tasks. In this paper, we propose Multi-level Disentangled Heterogeneous Graph Contrastive Learning method and learning disentangled HIN node representations in a self-supervised way. Specifically, we first design a tailored encoder to capture the latent factors and semantics of nodes in input HIN and learn their factorized representations. Then we propose a novel contrastive learning discrimination objective designed for disentangled HIN node representation learning. Extensive experiments conducted on various real-world datasets demonstrate the superiority of our method against state-of-the-art baselines.