Graph Multihead Attention Pooling with Self-Supervised Learning.

Area of Science:

• Graph Neural Networks (GNNs)
• Machine Learning
• Data Science

Background:

• Graph neural networks (GNNs) excel at graph-structured data tasks.
• Current GNNs often overlook graph pooling's importance for graph-level tasks.
• Effective pooling should retain local node features and global graph structure.

Purpose of the Study:

• To develop a hierarchical graph pooling method that minimizes information loss.
• To introduce GMAPS (Graph pooling based on Multihead Attention) for enhanced GNN performance.
• To improve graph classification and reconstruction tasks.

Main Methods:

• Proposed GMAPS, a hierarchical graph pooling method utilizing multihead attention.
• Nodes are arranged into a coarsened graph considering features and structural dependencies.
• A self-supervised mechanism maximizes mutual information between cluster and global representations.

Main Results:

• GMAPS demonstrated significant and consistent performance improvements over state-of-the-art baselines.
• The method achieved superior results on graph classification and reconstruction tasks.
• Experiments were conducted on six diverse benchmarks from biological and social domains.

Conclusions:

• GMAPS effectively compresses node features and graph structure into a coarsened graph.
• The proposed method enhances the expressiveness of cluster representations.
• GMAPS offers a robust solution for graph-level tasks in GNNs.