Harnessing collective structure knowledge in data augmentation for graph neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph neural networks (GNNs) excel at graph representation learning via message passing.
• Current GNNs often overlook crucial node and graph structural information, limiting their expressive power.
• Existing graph data augmentation methods struggle to scale with multiple structure features.

Purpose of the Study:

• To propose a novel approach, the collective structure knowledge-augmented graph neural network (CoS-GNN).
• To enable GNNs to leverage a diverse set of node- and graph-level structure features.
• To improve the modeling of structural knowledge in GNNs for enhanced graph representations.

Main Methods:

• Introduced a new message passing method within CoS-GNN.
• Integrated diverse node- and graph-level structure features with original node attributes.
• Augmented graphs to incorporate collective structural knowledge.

Main Results:

• CoS-GNN significantly enhances structural knowledge modeling at both node and graph levels.
• Achieved substantially improved graph representations compared to existing methods.
• Outperformed state-of-the-art models in graph classification, anomaly detection, and out-of-distribution generalization.

Conclusions:

• CoS-GNN effectively addresses the limitations of traditional GNNs by incorporating collective structure knowledge.
• The proposed method offers a scalable and powerful approach for advanced graph representation learning.
• Demonstrated superior performance across various graph-level learning tasks.