SP-GNN: Learning structure and position information from graphs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph Neural Networks (GNNs) are effective for graph data but often lack expressive power.
• Existing GNNs struggle to capture unique structural and positional information inherent in graph data.

Purpose of the Study:

• To introduce Structure- and Position-aware Graph Neural Networks (SP-GNN) to enhance GNNs' expressive capabilities.
• To develop methods for analyzing and leveraging structural and positional information in GNN tasks.

Main Methods:

• Proposed SP-GNN architecture incorporating a proximity-aware position encoder and a scalable structure encoder.
• Developed awareness scores to guide feature fusion strategies for improved GNN performance.
• Conducted extensive experiments on various graph datasets.

Main Results:

• SP-GNN demonstrated significant improvements in graph classification tasks compared to existing GNN models.
• The proposed encoders effectively capture and utilize structural and positional graph properties.
• Awareness scores provided insights into task-specific information requirements.

Conclusions:

• SP-GNN offers a generic and expressive framework for learning from graph data.
• Incorporating structural and positional awareness enhances GNN performance on downstream tasks.
• The SP-GNN approach provides a valuable tool for analyzing and optimizing GNNs.