Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cluster-guided Contrastive Class-imbalanced Graph Classification

Wei Ju, Zhengyang Mao, Siyu Yi, Yifang Qin, Yiyang Gu, Zhiping Xiao, Jianhao Shen, Ziyue Qiao, Ming Zhang

TL;DR

This work tackles class-imbalanced graph classification by introducing C$^3$GNN, which blends adaptive clustering, subclass-level Mixup, and hierarchical supervised contrastive learning to balance learning across majority and minority classes. By decomposing majority classes into semantically coherent subclasses with comparable sizes to minority classes and enriching them via Mixup, the method enables balanced, fine-grained representation learning across a hierarchy of class and subclass labels. The approach demonstrates superior performance over diverse baselines on six Benchmarks, with ablations validating the contribution of each component and analyses confirming balanced, informative feature distributions. The proposed framework offers a practical pathway to robust graph classification in real-world imbalanced domains, leveraging hierarchical structure to prevent minority-class overfitting while preserving majority-class richness.

Abstract

This paper studies the problem of class-imbalanced graph classification, which aims at effectively classifying the graph categories in scenarios with imbalanced class distributions. While graph neural networks (GNNs) have achieved remarkable success, their modeling ability on imbalanced graph-structured data remains suboptimal, which typically leads to predictions biased towards the majority classes. On the other hand, existing class-imbalanced learning methods in vision may overlook the rich graph semantic substructures of the majority classes and excessively emphasize learning from the minority classes. To address these challenges, we propose a simple yet powerful approach called C$^3$GNN that integrates the idea of clustering into contrastive learning to enhance class-imbalanced graph classification. Technically, C$^3$GNN clusters graphs from each majority class into multiple subclasses, with sizes comparable to the minority class, mitigating class imbalance. It also employs the Mixup technique to generate synthetic samples, enriching the semantic diversity of each subclass. Furthermore, supervised contrastive learning is used to hierarchically learn effective graph representations, enabling the model to thoroughly explore semantic substructures in majority classes while avoiding excessive focus on minority classes. Extensive experiments on real-world graph benchmark datasets verify the superior performance of our proposed method against competitive baselines.

Cluster-guided Contrastive Class-imbalanced Graph Classification

TL;DR

This work tackles class-imbalanced graph classification by introducing CGNN, which blends adaptive clustering, subclass-level Mixup, and hierarchical supervised contrastive learning to balance learning across majority and minority classes. By decomposing majority classes into semantically coherent subclasses with comparable sizes to minority classes and enriching them via Mixup, the method enables balanced, fine-grained representation learning across a hierarchy of class and subclass labels. The approach demonstrates superior performance over diverse baselines on six Benchmarks, with ablations validating the contribution of each component and analyses confirming balanced, informative feature distributions. The proposed framework offers a practical pathway to robust graph classification in real-world imbalanced domains, leveraging hierarchical structure to prevent minority-class overfitting while preserving majority-class richness.

Abstract

This paper studies the problem of class-imbalanced graph classification, which aims at effectively classifying the graph categories in scenarios with imbalanced class distributions. While graph neural networks (GNNs) have achieved remarkable success, their modeling ability on imbalanced graph-structured data remains suboptimal, which typically leads to predictions biased towards the majority classes. On the other hand, existing class-imbalanced learning methods in vision may overlook the rich graph semantic substructures of the majority classes and excessively emphasize learning from the minority classes. To address these challenges, we propose a simple yet powerful approach called CGNN that integrates the idea of clustering into contrastive learning to enhance class-imbalanced graph classification. Technically, CGNN clusters graphs from each majority class into multiple subclasses, with sizes comparable to the minority class, mitigating class imbalance. It also employs the Mixup technique to generate synthetic samples, enriching the semantic diversity of each subclass. Furthermore, supervised contrastive learning is used to hierarchically learn effective graph representations, enabling the model to thoroughly explore semantic substructures in majority classes while avoiding excessive focus on minority classes. Extensive experiments on real-world graph benchmark datasets verify the superior performance of our proposed method against competitive baselines.

Paper Structure

This paper contains 17 sections, 8 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Definition & Preliminaries
  3. Methodology
  4. Adaptive Clustering for Subclass-balancing
  5. Subclass Mixup Interpolation
  6. Hierarchical Subclass Contrastive Learning
  7. Computational Complexity Analysis
  8. Experiment
  9. Experimental Settings
  10. Experimental results
  11. Ablation study
  12. Hyper-parameter Sensitivity
  13. Analysis of Feature Distribution
  14. Visualization of Clustering Results
  15. Related Work
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Illustration of the proposed framework C$^3$GNN.
  • Figure 2: Sensitivity analysis of cluster size.
  • Figure 3: Feature distance and distribution curve of classes and subclasses on MNIST with imbalance factor 50.
  • Figure 4: Visualization of clustering results. We visualize several graphs in the MNIST dataset with their corresponding ground truth classes and resulting subclasses. Note that graphs (a) to (d) belong to class 3, where (a) and (b) belong to a different subclass from (c) and (d). On the other hand, graphs (e) to (h) are grouped under class 7, where (e) and (f) belong to a different subclass from (g) and (h).