Research map / Graph learning / Graph self-supervised learning
IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces IGL-Bench, an open-sourced unified benchmark evaluating 24 imbalanced graph learning algorithms on 17 node- and graph-level datasets for class- and topology-imbalance, assessing effectiveness, robustness, and efficiency.
Problem
Imbalanced graph learning (IGL) research lacks consistent datasets, splitting protocols, experimental settings, and metrics, so reported algorithm results are not comparable across studies.
Approach
A unified benchmark that fixes data processing and splits (1:1:8), controls imbalance ratios for class- and topology-imbalance (local degree and global reaching/squashing, graph size), and jointly measures accuracy, robustness across imbalance levels, representation boundaries (Silhouette/t-SNE), and time/memory cost for many IGL methods.
Key results
- GraphSHA reaches 91.55% accuracy on Photo under class-imbalance, and TOPOAUC reaches 83.34% on Cora and 69.03% on CiteSeer.
- Cold Brew achieves the best results on Chameleon (59.16%), Squirrel (43.04%), and Actor (33.01%) under local topology-imbalance.
- Nearly half of node-level IGL algorithms run out of memory on ogbn-arXiv, and most graph-level algorithms fail on ogbg-molhiv for efficiency reasons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use resampling-based class-imbalance methods instead of reweighting or data augmentation | Node-level class-imbalance algorithms · Node classification across datasets | accuracy robustness as imbalance increases | Helped |
| Use data augmentation rather than resampling | Node-level class-imbalance algorithms · Six of nine node datasets | accuracy | Helped |
| Apply TOPOAUC | Node-level class-imbalanced classification · Cora | accuracy: 76.36±0.13 -> 83.34±0.31 | Helped |
| Apply Cold Brew | Node-level local topology-imbalance classification · Actor | accuracy: 21.55±1.74 -> 33.01±0.19 | Helped |
| Apply PASTEL | Node-level global topology-imbalance classification · Chameleon | accuracy: 33.90±0.70 -> 47.12±2.82 | Helped |
| Apply TOPOAUC | Node-level class-imbalance classification · heterophilic datasets | accuracy | Hurt |
| Apply DataDec | Graph-level class-imbalance classification · Graph datasets | accuracy | Helped |
| Apply G²GNN | Graph-level class-imbalance classification · Binary versus multi-class datasets | accuracy | Mixed |
| Apply SOLT-GNN | Graph-level topology-imbalance classification · Five datasets | accuracy | Helped |
| Increase the imbalance ratio | Node-level IGL algorithms · Node-level classification | accuracy | Hurt |
| Apply neighbor augmentation | Node-level local topology-imbalance methods · Homophilic datasets | accuracy | Helped |
| Apply neighbor augmentation | Node-level local topology-imbalance methods · Heterophilic datasets such as Actor | accuracy robustness across imbalance levels | Hurt |
Surprising findings
- Most IGL algorithms do not consistently beat vanilla GCN/GIN; TopoImb fails to outperform its backbone consistently.
- Many node-level IGL algorithms fail or run out of memory on the large-scale ogbn-arXiv.
Limitations
- Many algorithms run out of memory or time out on large graphs such as ogbn-arXiv and ogbg-molhiv.
- Class- and topology-imbalance interactions are not resolved; no unified algorithm is proposed.
Experimental setup
Datasets: Cora CiteSeer PubMed Computers Photo ogbn-arXiv Chameleon Squirrel Actor PTC-MR FRANKENSTEIN PROTEINS
Benchmarks: Cora CiteSeer PubMed Computers Photo ogbn-arXiv Chameleon Squirrel Actor PTC-MR FRANKENSTEIN PROTEINS
Metrics: Accuracy balanced accuracy Macro-F1 AUC-ROC Silhouette score training time peak GPU memory
Baselines: GCN GIN DRGCN DPGNN ImGAGN GraphSMOTE GraphENS GraphMixup LTE4G TAM TOPOAUC GraphSHA
Builds on: GraphSMOTE GraphENS TAM TOPOAUC PASTEL
Closest papers by meaning
- Class-Imbalanced Graph Learning without Class Rebalancing (ICML 2024 · similarity 0.89)
- NoisyGL: A Comprehensive Benchmark for Graph Neural Networks under Label Noise (NeurIPS 2024 · similarity 0.88)
- Cluster-guided Contrastive Class-imbalanced Graph Classification (AAAI 2025 · similarity 0.88)
- No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets (ICML 2025 · similarity 0.88)
- GLBench: A Comprehensive Benchmark for Graph with Large Language Models (NeurIPS 2024 · similarity 0.87)
- On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.