HGBER: Heterogeneous Graph Neural Network With Bidirectional Encoding Representation.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Heterogeneous graphs are common in real-world applications.
• Heterogeneous Graph Neural Networks (HGNNs) effectively process these graphs.
• Existing HGNNs often oversimplify relationships between meta-paths.

Purpose of the Study:

• To propose a novel unsupervised framework, Heterogeneous Graph neural network with bidirectional encoding representation (HGBER).
• To address limitations in existing HGNNs by capturing more general and complex relationships between meta-paths.
• To learn comprehensive and structure-preserving node representations.

Main Methods:

• Contrastive forward encoding to extract meta-specific node representations.
• Reversed encoding for a degradation process from final to meta-specific representations.
• A self-training module for iterative optimization and structure preservation.

Main Results:

• HGBER outperforms state-of-the-art HGNN baselines.
• Accuracy improvements range from 0.8% to 8.4% across datasets.
• Demonstrated effectiveness on various downstream tasks.

Conclusions:

• HGBER offers a superior approach to learning node representations in heterogeneous graphs.
• The bidirectional encoding and self-training modules effectively capture complex relationships.
• The framework shows significant potential for advancing HGNN research and applications.