A Self-Supervised Heterogeneous Graph Attention Model Based on Adaptable Step-Size Metapaths.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Analysis

Background:

• Heterogeneous graph neural networks (HGNNs) are crucial for modeling complex real-world networks.
• Existing HGNNs often require predefined metapaths and struggle with limited labeled data.
• Current methods lack metapath sequence modeling and adaptive feature extraction.

Purpose of the Study:

• To propose a self-supervised heterogeneous graph attention model (HGAM) using adaptable step-size metapaths.
• To overcome limitations of predefined metapaths and address data scarcity in HGNNs.
• To enhance representation learning by adaptively capturing important metapaths and integrating global information.

Main Methods:

• Developed an adaptable step-size metapaths module for HGAM, considering attention weights and trends across different step sizes.
• Implemented a dual contrastive learning strategy for self-supervised learning, contrasting high-order meta-graphs with nodes and preserving local structure.
• Evaluated HGAM on node classification, clustering, and link prediction tasks using real-world datasets.

Main Results:

• HGAM adaptively captures important step-size metapaths, expanding the model's receptive field and integrating global information.
• The dual contrastive learning strategy effectively addresses labeled data scarcity.
• Achieved superior performance compared to state-of-the-art methods across all evaluated tasks.

Conclusions:

• HGAM offers a novel and effective approach to self-supervised learning on heterogeneous graphs.
• The adaptable step-size metapaths and dual contrastive learning significantly improve representation learning.
• HGAM demonstrates strong potential for diverse graph-based analytical tasks.