Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ECGN: A Cluster-Aware Approach to Graph Neural Networks for Imbalanced Classification

Bishal Thapaliya, Anh Nguyen, Yao Lu, Tian Xie, Igor Grudetskyi, Fudong Lin, Antonios Valkanas, Jingyu Liu, Deepayan Chakraborty, Bilel Fehri

TL;DR

The Enhanced Cluster-aware Graph Network (ECGN), a novel method that addresses cluster-specific training with synthetic node generation by integrating cluster-specific training with synthetic node generation, and enhances the quality of the resulting node embeddings.

Abstract

Classifying nodes in a graph is a common problem. The ideal classifier must adapt to any imbalances in the class distribution. It must also use information in the clustering structure of real-world graphs. Existing Graph Neural Networks (GNNs) have not addressed both problems together. We propose the Enhanced Cluster-aware Graph Network (ECGN), a novel method that addresses these issues by integrating cluster-specific training with synthetic node generation. Unlike traditional GNNs that apply the same node update process for all nodes, ECGN learns different aggregations for different clusters. We also use the clusters to generate new minority-class nodes in a way that helps clarify the inter-class decision boundary. By combining cluster-aware embeddings with a global integration step, ECGN enhances the quality of the resulting node embeddings. Our method works with any underlying GNN and any cluster generation technique. Experimental results show that ECGN consistently outperforms its closest competitors by up to 11% on some widely studied benchmark datasets.

ECGN: A Cluster-Aware Approach to Graph Neural Networks for Imbalanced Classification

TL;DR

The Enhanced Cluster-aware Graph Network (ECGN), a novel method that addresses cluster-specific training with synthetic node generation by integrating cluster-specific training with synthetic node generation, and enhances the quality of the resulting node embeddings.

Abstract

Classifying nodes in a graph is a common problem. The ideal classifier must adapt to any imbalances in the class distribution. It must also use information in the clustering structure of real-world graphs. Existing Graph Neural Networks (GNNs) have not addressed both problems together. We propose the Enhanced Cluster-aware Graph Network (ECGN), a novel method that addresses these issues by integrating cluster-specific training with synthetic node generation. Unlike traditional GNNs that apply the same node update process for all nodes, ECGN learns different aggregations for different clusters. We also use the clusters to generate new minority-class nodes in a way that helps clarify the inter-class decision boundary. By combining cluster-aware embeddings with a global integration step, ECGN enhances the quality of the resulting node embeddings. Our method works with any underlying GNN and any cluster generation technique. Experimental results show that ECGN consistently outperforms its closest competitors by up to 11% on some widely studied benchmark datasets.

Paper Structure

This paper contains 29 sections, 6 equations, 3 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Class Imbalance Learning
  4. Graph Neural Networks
  5. Proposed Algorithm
  6. Architecture of ECGN
  7. Cluster-Aware SMOTE
  8. Experiments
  9. Results
  10. Direct Inference from Subclusters: Why Do We Need Global Integration?
  11. Reusing GNN Weights from Pre-training in Global Integration
  12. Selecting the Number of Clusters
  13. Conclusion and Limitations
  14. Appendix
  15. Details of Baselines and Experimental Datasets
  16. ...and 14 more sections

Figures (3)

  • Figure 1: Working framework of ECGN architecture. We perform clustering on the nodes initially, train the sub-clusters independently in parallel, update the node embeddings to the original graph followed by Cluster-Aware SMOTE step and finally global integration.
  • Figure 2: Effect of Number of Clusters on F1-Score for Various Datasets
  • Figure 3: Visualizations of the original and clustered versions of the Cora and Citeseer datasets. The left column shows the original datasets, while the right column shows the datasets divided into three clusters using METIS clustering.