[Paper Review] CommPOOL: An Interpretable Graph Pooling Framework for Hierarchical Graph Representation Learning

Haoteng Tang, Guixiang Ma+3 more|arXiv (Cornell University)|Dec 10, 2020

Advanced Graph Neural Networks56 references4 citations

TL;DR

This paper proposes CommPOOL, a novel interpretable graph pooling framework that captures and preserves hierarchical community structures in graphs through unsupervised community detection using PAM clustering. By pooling nodes based on community medoids and aggregating structural and feature information, CommPOOL achieves state-of-the-art performance in graph classification while enabling transparent, human-interpretable pooling operations across five benchmark datasets and one synthetic dataset.

ABSTRACT

Recent years have witnessed the emergence and flourishing of hierarchical graph pooling neural networks (HGPNNs) which are effective graph representation learning approaches for graph level tasks such as graph classification. However, current HGPNNs do not take full advantage of the graph's intrinsic structures (e.g., community structure). Moreover, the pooling operations in existing HGPNNs are difficult to be interpreted. In this paper, we propose a new interpretable graph pooling framework - CommPOOL, that can capture and preserve the hierarchical community structure of graphs in the graph representation learning process. Specifically, the proposed community pooling mechanism in CommPOOL utilizes an unsupervised approach for capturing the inherent community structure of graphs in an interpretable manner. CommPOOL is a general and flexible framework for hierarchical graph representation learning that can further facilitate various graph-level tasks. Evaluations on five public benchmark datasets and one synthetic dataset demonstrate the superior performance of CommPOOL in graph representation learning for graph classification compared to the state-of-the-art baseline methods, and its effectiveness in capturing and preserving the community structure of graphs.

Motivation & Objective

• To address the lack of interpretability in existing hierarchical graph pooling networks (HGPNNs), especially in how pooling operations are performed.
• To preserve the intrinsic hierarchical community structure of graphs during graph representation learning, which is often ignored in standard GNNs.
• To develop a general, flexible, and interpretable framework for graph-level representation learning that enhances performance on graph classification tasks.
• To demonstrate that community structure is a critical factor in graph classification by comparing community-based pooling with semi-random pooling.

Proposed method

• CommPOOL employs an unsupervised community detection method using Partitioning Around Medoids (PAM) to identify natural communities in graphs based on node feature similarity.
• The framework performs hierarchical pooling by selecting the medoid node of each detected community as the representative node for that community.
• Node features and structural information from all members of a community are aggregated into the medoid node to preserve local and community-level representations.
• The pooling process is inherently interpretable because the community structure is explicitly detected and visualized, allowing users to trace how nodes are grouped and downsampled.
• The method is integrated into a GNN-based architecture that jointly learns local node features and hierarchical community-level representations.
• The framework is evaluated using a GNN backbone with multiple pooling layers, enabling end-to-end training for graph classification.

Experimental results

Research questions

• RQ1 How can graph hierarchical structures be captured and preserved in a way that is both effective and interpretable?
• RQ2 Can community-based pooling improve graph classification performance compared to non-community-aware pooling strategies?
• RQ3 To what extent does preserving the true community structure impact model performance and robustness?
• RQ4 Is the pooling operation transparent and interpretable, allowing users to understand which nodes are grouped and why?

Key findings

• CommPOOL achieves state-of-the-art performance on five public graph classification benchmarks, outperforming existing SOTA methods.
• On the PROTEINS dataset, CommPOOL achieved 74.74% accuracy, significantly higher than the semi-random pooling baseline of 64.90%.
• On the BZR dataset, CommPOOL reached 86.00% accuracy, compared to 81.50% for semi-random pooling, demonstrating the importance of proper community structure preservation.
• The synthetic simulation dataset showed that CommPOOL achieved 80.14% accuracy, compared to 70.34% for semi-random pooling, further validating the role of community structure.
• Visualizations confirm that CommPOOL successfully captures meaningful community structures, while misclassification occurs when communities are poorly defined.
• The ablation study shows that community-based pooling is significantly more effective than semi-random partitioning, even when the latter maintains some hierarchical structure.