CCP-GNN: Competitive Covariance Pooling for Improving Graph Neural Networks

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Science

Background:

• Graph neural networks (GNNs) are crucial for graph classification, with global pooling methods summarizing node features to create graph representations.
• Existing global average pooling (GAP) methods in GNNs overlook node specificity and rich feature statistics, limiting classification performance.

Purpose of the Study:

• To propose a novel Competitive Covariance Pooling (CCP) method for GNNs to improve graph classification performance.
• To address limitations in existing pooling methods by considering node specificity and feature statistics.

Main Methods:

• CCP generates node-level second-order representations to capture rich statistics within node features.
• A competitive attention module identifies key nodes by learning node weights.
• Node-level representations and weights are aggregated into a covariance representation, refined by iterative matrix normalization.

Main Results:

• The proposed CCP, integrated into GNNs as CCP-GNN, significantly improves graph classification performance.
• Experiments on seven graph-level benchmarks demonstrate CCP-GNN's superiority or competitiveness against state-of-the-art methods.
• CCP enhances GNN performance with minimal computational overhead.

Conclusions:

• CCP offers an effective approach to graph representation learning by leveraging node-specific statistics and identifying key nodes.
• CCP-GNN provides a flexible and powerful enhancement for various GNN architectures in graph classification tasks.
• The method shows promising results, advancing the capabilities of GNNs in analyzing complex graph structures.