A multi-view contrastive learning for heterogeneous network embedding.

Area of Science:

• Graph representation learning
• Machine learning on networks
• Artificial intelligence

Background:

• Homogeneous graph contrastive learning learns node representations but struggles with heterogeneous information networks (HINs).
• Augmenting HINs and designing suitable pretext tasks for contrastive learning remain challenging.
• Existing debiasing techniques are insufficient for graph contrastive learning, particularly on HINs.

Purpose of the Study:

• To propose a novel multi-view heterogeneous graph contrastive learning framework (MCL).
• To address challenges in HIN augmentation, pretext task design, and sampling bias mitigation.
• To enhance discriminative node representations in HINs.

Main Methods:

• Utilizing metapaths to generate multiple subgraphs (views) for data augmentation.
• Implementing a novel pretext task to maximize coherence between metapath-induced views.
• Employing a positive sampling strategy considering semantics and structures to alleviate sampling bias.

Main Results:

• The proposed MCL framework demonstrates superior performance compared to state-of-the-art baselines.
• MCL achieves competitive results, even surpassing supervised methods in certain scenarios.
• Experiments conducted on five real-world benchmark datasets validate the framework's effectiveness.

Conclusions:

• The multi-view heterogeneous graph contrastive learning framework (MCL) effectively addresses key challenges in HIN representation learning.
• MCL provides a robust approach for learning discriminative node embeddings by leveraging metapath-induced views and mitigating sampling bias.
• The proposed method offers a significant advancement in contrastive learning for complex, heterogeneous graph data.