CommPOOL: An interpretable graph pooling framework for hierarchical graph representation learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Hierarchical graph pooling neural networks (HGPNNs) are advanced methods for graph representation learning.
• Existing HGPNNs often overlook intrinsic graph structures like community structure.
• Interpretability of pooling operations in current HGPNNs is limited.

Purpose of the Study:

• To propose CommPOOL, an interpretable graph pooling framework.
• To capture and preserve hierarchical community structures within graphs.
• To enhance graph representation learning for graph-level tasks.

Main Methods:

• Developed a novel community pooling mechanism within CommPOOL.
• Employed an unsupervised approach to identify inherent graph community structures.
• Designed CommPOOL as a general and flexible framework for hierarchical graph representation learning.

Main Results:

• CommPOOL demonstrated superior performance in graph classification tasks.
• Achieved state-of-the-art results compared to existing baseline methods.
• Effectively captured and preserved the community structure of graphs across evaluations.

Conclusions:

• CommPOOL offers an interpretable and effective approach to graph representation learning.
• The framework successfully leverages hierarchical community structures for improved performance.
• CommPOOL shows significant potential for advancing graph-level tasks.