CommPOOL: An Interpretable Graph Pooling Framework for Hierarchical Graph Representation Learning

Haoteng Tang, Guixiang Ma, Lifang He, Heng Huang, Liang Zhan

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.

Introduction

Graph Neural Networks (GNNs) have revolutionized how we analyze complex data, from social networks to chemical molecules. However, most GNNs suffer from a "forest for the trees" problem: they excel at understanding individual nodes and their immediate neighbors but often struggle to capture broader, hierarchical structures—like communities—that define a graph's true nature. When classifying an entire graph (e.g., determining molecular toxicity), standard "global pooling" strategies like simple summation or averaging are often destructive, ignoring the hierarchical patterns and functional clusters common in real-world networks.

Existing Hierarchical Graph Pooling (HGP) methods attempt to address this by gradually downsampling the graph. However, many of these are "black boxes," using trainable parameters to group nodes in ways that are difficult for humans to interpret. CommPOOL addresses these challenges by providing a framework that is both high-performing and inherently transparent.

Methodology: The CommPOOL Framework

The CommPOOL framework is built around Embedding-Pooling (EP) modules that can be stacked to create a hierarchy. Each module consists of two main stages:

  1. GNN-based Node Embedding: CommPOOL utilizes a Variational Graph Auto-Encoder (VGAE). This unsupervised approach learns node features by reconstructing the graph's own structure, ensuring that embeddings reflect intrinsic topology rather than just fitting the final task label.
  2. Community Pooling: This stage employs an unsupervised clustering algorithm, Partitioning Around Medoids (PAM). It identifies "medoids" (representative center nodes) for a set number of communities and assigns every other node to its nearest medoid based on feature similarity. The features are then pooled into the medoid using a similarity-weighted sum, preserving the connections between communities in the down-scaled graph.
The CommPOOL architecture, illustrating the flow from node embedding via VGAE to community detection and pooling, followed by graph classification.

By using medoids—actual nodes—as community centers rather than virtual centroids, the pooling process remains grounded in the original data, significantly enhancing interpretability.

Key Results and Evaluation

The authors evaluated CommPOOL on five public benchmark datasets (PROTEINS, Synthie, FRANKENSTEIN, BZR, AIDS) and a custom synthetic dataset designed with ground-truth community labels.

CommPOOL consistently outperformed popular baselines like DIFFPOOL, SAG-POOL, and SortPool. Notably, on the BZR dataset, CommPOOL achieved 86.00% accuracy, a ~5% improvement over the best baseline. On the Synthie dataset, it reached 66.50%, significantly outperforming HGP-SL (54.00%).

Performance metrics on synthetic data showing high NMI scores (a) and visual comparisons of community detection results (b, c).

On the synthetic dataset, the model achieved a Normalized Mutual Information (NMI) score of 0.952, indicating that the communities discovered by the model were nearly identical to the ground truth. Visualizations further demonstrate that correct identification of community structures correlates strongly with accurate classification, while "messy" detection leads to errors.

Discussion and Future Work

CommPOOL demonstrates that communities are a natural and effective unit for representing graph hierarchies, aligning well with domain knowledge like functional groups in chemistry. By utilizing unsupervised methods for structure capture, the model preserves the "true" shape of the graph.

While powerful, the framework currently relies on fixed hyperparameters for the number of communities and pooling layers. Future work could involve learning these parameters dynamically. Additionally, the authors suggest extending the framework to graph regression tasks and further optimizing the computational complexity of clustering algorithms like PAM for extremely large-scale graphs.

Summary updated

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.
BibTeX
@inproceedings{tang2021commpool,
  author = {Haoteng Tang and Guixiang Ma and Lifang He and Heng Huang and Liang Zhan},
  title = {{CommPOOL: An Interpretable Graph Pooling Framework for Hierarchical Graph Representation Learning}},
  booktitle = {Neural Networks},
  year = {2021},
  url = {https://openreview.net/forum?id=w2NZ7tlPG30},
}
Figure