Topology-Aware Graph Pooling Networks.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Computer Vision

Background:

• Pooling operations are vital in computer vision and natural language processing.
• Graph data presents unique challenges for pooling due to undefined locality.
• Existing methods often fail to incorporate graph topology effectively.

Purpose of the Study:

• To propose a novel pooling layer for graph data that explicitly considers graph topology.
• To improve node selection in graphs by integrating local and global structural information.
• To enhance performance on graph classification tasks.

Main Methods:

• Introduction of the topology-aware pooling (TAP) layer.
• A two-stage voting process involving local and global node importance scoring.
• Incorporation of graph connectivity into ranking score computation.

Main Results:

• The TAP layer explicitly considers graph topology during node selection.
• Local voting attends to neighboring nodes, while global voting considers the entire graph.
• Combining local and global scores, along with a connectivity term, yields improved performance.

Conclusions:

• The proposed TAP layer effectively addresses the challenge of pooling on graph data.
• TAP achieves superior performance on graph classification tasks compared to previous methods.
• The method enhances graph representation learning by leveraging topological information.