Not all edges are peers: Accurate structure-aware graph pooling networks

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Mining

Background:

• Graph Neural Networks (GNNs) excel in graph-related tasks, with pooling operators crucial for graph-level representation learning in graph classification.
• Existing pooling methods often discard edges and nodes, leading to information loss and incomplete structural utilization.
• Current approaches may overemphasize certain substructures, neglecting others and hindering comprehensive graph representation.

Purpose of the Study:

• To introduce a novel pooling operator, Accurate Structure-Aware Graph Pooling (ASPool), for enhanced graph-level representation learning in GNNs.
• To address limitations of existing pooling methods by preserving graph connectivity and utilizing structural information more effectively.
• To improve the performance of GNNs in graph classification tasks through a more accurate and comprehensive pooling strategy.

Main Methods:

• ASPool adaptively retains a subset of edges to refine graph structure, treating edges as non-peers rather than simple node connectors.
• A selection strategy is implemented to preserve graph connectivity by considering both top-ranked nodes and retained edges.
• A two-stage calculation process ensures sampled nodes are distributed throughout the graph, capturing diverse structural information.

Main Results:

• ASPool demonstrates superior performance compared to state-of-the-art graph representation learning methods.
• Experiments conducted on 9 widely used benchmarks validate the effectiveness of the proposed ASPool operator.
• The method successfully addresses information loss and connectivity issues present in previous pooling techniques.

Conclusions:

• ASPool offers a significant advancement in graph pooling for GNNs, enabling more robust graph-level representations.
• The proposed method enhances graph classification accuracy by preserving crucial structural information and connectivity.
• ASPool provides a versatile and effective solution for integrating into various GNN architectures for improved performance.