Cross-view self-supervised heterogeneous graph representation learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Heterogeneous graph neural networks (HGNNs) struggle with integrating multi-view information, limiting their effectiveness on complex data.
• Existing methods often fail to fully exploit the rich structural and semantic information present in heterogeneous graphs.

Purpose of the Study:

• To develop an improved graph-level cross-attention mechanism for HGNNs to enhance multi-view integration.
• To boost the expressiveness and performance of models on complex, multi-view heterogeneous network data.

Main Methods:

• Incorporated random walks, Katz index, and Transformers to capture higher-order semantic relationships within meta-path views.
• Utilized network decomposition and attention mechanisms for node context extraction in network schema views.
• Employed an improved graph-level cross-attention for adaptive feature fusion across views and a contrastive loss function for sample selection.

Main Results:

• The proposed self-supervised model demonstrated superior performance in node classification and clustering tasks.
• The enhanced cross-attention mechanism effectively fused features from multiple views, improving model expressiveness.
• The contrastive loss function enhanced model robustness by leveraging local and global node centrality.

Conclusions:

• The developed graph-level cross-attention mechanism significantly improves multi-view integration in HGNNs.
• The self-supervised approach offers an effective solution for leveraging complex heterogeneous graph data.
• The method shows strong potential for applications requiring advanced node classification and clustering.