---

IGL-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning (ICLR 2025 Spotlight🌟) [PDF]

IGL-Bench is a comprehensive benchmark for Imbalanced Graph Learning (IGL) based on PyTorch and PyTorch Geometric. We embark on 16 diverse graph datasets and 24 distinct IGL algorithms with uniform data processing and splitting strategies.

📔 Overview of the IGL-Bench

IGL-Bench serves as the first open-sourced benchmark for graph-specific imbalanced learning to the best of our knowledge. IGL-Bench encompases 24 state-of-the-art IGL algorithms and 16 diverse graph datasets covering node-level and graph-level tasks, addressing class- and topology-imbalance issues, while also adopting consistent data processing and splitting approaches for fair comparisons over multiple metrics with different investigation focus. Through benchmarking the existing IGL algorithms for effectiveness, robustness, and efficiency, we make the following contributions:

• First Comprehensive IGL Benchmark. IGL-Bench enables a fair and unified comparison among 19 state-of-the-art node-level and 5 graph-level IGL algorithms by unifying the experimental settings across 16 graph datasets of diverse characteristics, providing a comprehensive understanding of the class-imbalance and topology-imbalance problems in IGL for the first time.
• Multi-faceted Evaluation and Analysis. We conduct a systematic analysis of IGL methods from various dimensions, including effectiveness, efficiency, and complexity. Based on the results of extensive experiments, we uncover both the potential advantages and limitations of current IGL algorithms, providing valuable insights to guide future research endeavors.
• Open-sourced Package. To facilitate future IGL research, we develop an open-sourced benchmark package for public access. Users can evaluate their algorithms or datasets with less effort.

📦 Installation

Follow the steps below to install and configure IGL-Bench properly for your local environment.

Environment Requirements

Please ensure your Python environment meets following dependencies:

Dependency	Version (≥)
Python	3.8.12
PyTorch	1.9.1
PyTorch-Geometric	2.0.1
DGL	1.1
scipy	1.9

Recommendation: Use your specific Python virtual environment (e.g., via conda or venv) or use this Dockerfilewe provide to build a Docker image.

[Recommended] Installation Steps

Clone the repository and install required dependencies:

# Clone the repository
git clone https://github.com/RingBDStack/IGL-Bench.git
cd IGL-Bench

# Install dependencies
pip install -r requirements.txt

[Optional] Development Mode Installation

If you want to install the package for your local development, use:

pip install -e .

🚀 Quick Start

The following demonstrates how to quickly run algorithms from IGL-Bench for both node-level and graph-level classification tasks under various imbalance settings.

You can also refer to our runnable example : 📓 demo.ipynb

🔹 Node-Level Task (e.g., PASTEL on Cora)

Step 1: Import Package

import IGL_Bench as igl

Step 2: Initialize Dataset

Dataset = igl.dataset.Dataset(
    task="node",
    data_name="Cora",
    imb_type="topo_global",
    imb_level="high"
)
dataset = Dataset.load_dataset()

Step 3: Load Configuration

config = igl.config.load_conf(
    task="node",
    imbtype="topo_global",
    algorithm="PASTEL"
)

Step 4: Run Solver

solver = igl.manage.Manager(config, dataset)
solver.run(num_runs=5)

🔸 Graph-Level Task (e.g., G2GNN on D&D)

Step 1: Import Package

import IGL_Bench as igl

Step 2: Initialize Dataset

Dataset = igl.dataset.Dataset(
    task="graph",
    data_name="D&D",
    imb_type="class",
    imb_level="low"
)
dataset = Dataset.load_dataset()

Step 3: Load Configuration

config = igl.config.load_conf(
    task="graph",
    imbtype="class",
    algorithm="G2GNN"
)

Step 4: Run Solver

solver = igl.manage.Manager(config, dataset)
solver.run(num_runs=10)

⚙️ Configuration Argument Options

You can flexibly configure algorithms using the Dataset(...) and load_conf(...) API with the following parameters:

task

Specifies the type of task:

'node' or 'graph'

data_name

Node-level datasets:

'Cora', 'CiteSeer', 'PubMed', 'Computers', 'Photo', 'ogbn-arxiv', 'Chameleon', 'Squirrel', 'Actor'

Graph-level datasets:

'PTC-MR', 'FRANKENSTEIN', 'PROTEINS', 'D&D', 'IMDB-B', 'REDDIT-B', 'COLLAB', 'ogbg-molhiv'

imbtype

Imbalance type (depends on task):

• For node-level:

  'class', 'topo_local', 'topo_global'
  

• For graph-level:

  'class', 'topology'
  

imb_level

Controls the severity of imbalance:

'low', 'mid', 'high'

algorithm

The algorithm to run. Supported algorithms include:

• Node-level class imbalance:

  DRGCN, DPGNN, ImGAGN, GraphSMOTE, GraphENS, GraphMixup, LTE4G, TAM, TOPOAUC, GraphSHA
  

• Node-level topology local imbalance:

  DEMONet, TailGNN, COLDBREW, LTE4G, meta-tail2vec, GRAPHPATCHER, RAWLSGCN
  

• Node-level topology global imbalance:

  ReNode, TAM, PASTEL, TOPOAUC, HyperIMBA
  

• Graph-level class imbalance:

  G2GNN, TopoImb, DataDec, ImGKB
  

• Graph-level topology imbalance:

  TopoImb, SOLTGNN
  

You can also manually modify the configuration files located in config/ for fine-grained control over hyperparameters.

🚧 Build Your Own IGL Algorithms

IGL-Bench is designed with modularity and extensibility. Due to the unified solver interface, users can seamlessly integrate their custom algorithms into our benchmarking pipeline with minimal redundant codes.

We provide a step-by-step tutorial to guide you through the process: 📓 Build_Your_Own_IGL.ipynb

This tutorial covers:

• ✅ Using the pre-split datasets (with train_mask, val_mask, test_mask)
• ✅ Understanding the Dataset and Config structure
• ✅ Adding your own algorithm in a few easy steps

Your custom solver will automatically be discovered and benchmarked by the framework—just like any built-in baseline.

🧩 Algorithm References

We have implemented the following IGL algorithms in the IGL-Bench:

Algorithm	Conference/Journal	Imbalance Type	TASK	Paper	Code
DRGCN	IJCAI 2020	Class-Imbalance	NC	Multi-Class Imbalanced Graph Convolutional Network Learning	Link
DPGNN	arXiv 2020	Class-Imbalance	NC	Distance-wise Prototypical Graph Neural Network for Imbalanced Node Classification	Link
ImGAGN	SIGKDD 2021	Class-Imbalance	NC	ImGAGN: Imbalanced Network Embedding via Generative Adversarial Graph Networks	Link
GraphSMOTE	WSDM 2021	Class-Imbalance	NC	GraphSMOTE: Imbalanced Node Classification on Graphs with Graph Neural Networks	Link
GraphENS	ICLR 2021	Class-Imbalance	NC	Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification	Link
GraphMixup	ECML PKDD 2022	Class-Imbalance	NC	GraphMixup: Improving Class-Imbalanced Node Classification on Graphs by Self-supervised Context Prediction	Link
LTE4G	CIKM 2022	Class/Local Topology-Imbalance	NC	LTE4G: Long-Tail Experts for Graph Neural Networks	Link
TAM	ICML 2022	Class/Global Topology-Imbalance	NC	TAM: Topology-Aware Margin Loss for Class-Imbalanced Node Classification	Link
TOPOAUC	ACM MM 2022	Class/Global Topology-Imbalance	NC	A Unified Framework against Topology and Class Imbalance	Link
GraphSHA	SIGKDD 2023	Class-Imbalance	NC	GraphSHA: Synthesizing Harder Samples for Class-Imbalanced Node Classification	Link
DEMO-Net	SIGKDD 2019	Local Topology-Imbalance	NC	DEMO-Net: Degree-specific Graph Neural Networks for Node and Graph Classification	Link
meta-tail2vec	CIKM 2020	Local Topology-Imbalance	NC	Towards locality-aware meta-learning of tail node embeddings on networks	Link
Tail-GNN	SIGKDD 2021	Local Topology-Imbalance	NC	Tail-GNN: Tail-Node Graph Neural Networks	Link
Cold Brew	ICLR 2022	Local Topology-Imbalance	NC	Cold Brew: Distilling Graph Node Representations with Incomplete or Missing Neighborhoods	Link
RawlsGCN	WWW 2022	Local Topology-Imbalance	NC	RawlsGCN: Towards Rawlsian Difference Principle on Graph Convolutional Network	Link
GraphPatcher	NeuIPS 2023	Local Topology-Imbalance	NC	GRAPHPATCHER: Mitigating Degree Bias for Graph Neural Networks via Test-time Augmentation	Link
ReNode	NeurIPS 2021	Global Topology-Imbalance	NC	Topology-Imbalance Learning for Semi-Supervised Node Classification	Link
PASTEL	CIKM 2022	Global Topology-Imbalance	NC	Position-aware Structure Learning for Graph Topology-imbalance by Relieving Under-reaching and Over-squashing	Link
HyperIMBA	WWW 2023	Global Topology-Imbalance	NC	Hyperbolic Geometric Graph Representation Learning for Hierarchy-imbalance Node Classification	Link
G2GNN	CIKM 2022	Class-Imbalance	GC	Imbalanced Graph Classification via Graph-of-Graph Neural Networks	Link
TopoImb	LOG 2022	Class/Topology-Imbalance	NC/GC	TopoImb: Toward Topology-level Imbalance in Learning from Graphs	Link
DataDec	ICML 2023	Class-Imbalance	NC/GC	When Sparsity Meets Contrastive Models: Less Graph Data Can Bring Better Class-Balanced Representations	Link
ImGKB	ACM MM 2023	Class-Imbalance	GC	Where to Find Fascinating Inter-Graph Supervision: Imbalanced Graph Classification with Kernel Information Bottleneck	Link
SOLT-GNN	WWW 2022	Class-Imbalance	GC	On Size-Oriented Long-Tailed Graph Classification of Graph Neural Networks	Link

📚 Citation

If you find this repository helpful, please consider citing us. We welcome any discussions with [email protected].

@inproceedings{qin2025iglbench,
  title={{IGL}-Bench: Establishing the Comprehensive Benchmark for Imbalanced Graph Learning},
  author={Jiawen Qin and Haonan Yuan and Qingyun Sun and Lyujin Xu and Jiaqi Yuan and Pengfeng Huang and Zhaonan Wang and Xingcheng Fu and Hao Peng and Jianxin Li and Philip S. Yu},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}