IGNN

NeurIPS 25 Poster:

Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective.

🛠️ Installation

• Python ≥ 3.8.16
• See virtual environment setup scripts:
  • Development: ./.ci/install-dev.sh
  • Production: ./.ci/install.sh
• Dependency lists:
  • ./requirements-dev.txt
  • ./requirements.txt

🚀 Usage

Run each IGNN variant across all datasets using the following scripts on:

1. Public splits: ./scripts/00-best-racIGNN-public.sh
2. Our splits:

Variant	Script
c-IGNN	./scripts/01-best-cIGNN.sh
r-IGNN	./scripts/02-best-rIGNN.sh
a-IGNN	./scripts/03-best-aIGNN.sh

3. Results of the public and our splits are documentd in ./results/table_pub.csv and ./results/table_our.csv.

Important

Experimental setups for our reported results:

• [Setting 1] Tesla V100, with Python 3.9.15, PyTorch 2.0.1, and Cuda 11.7.

We observed performance discrepancies when using identical hyperparameters across different PyTorch/CUDA versions and GPU architectures, e.g., V100 (Setting 1) vs. RTX 3090 (Setting 2).

• [Setting 2] RTX 3090, with Python 3.8.16, PyTorch 2.1.2, and Cuda 12.1.

For instance, on Chameleon, the same config in ./scripts/01-best-cIGNN.sh yielded 50.79 ± 4.92 (Setting 1) vs. 47.53 ± 3.36 (Setting 2). Although SOTA performance can be achieved under all environments with proper tuning, optimal hyperparameters may differ across setups.

Perform hyperparameter searches for each variant using:

Split	Search Script
Ours	./scripts/00-search-ours-split.sh
Public	./scripts/00-search-public-split.sh

Variant	Script
c-IGNN	./scripts/01cignn_search.py
r-IGNN	./scripts02rignn_search.py
a-IGNN	./scripts/03aignn_search.py

Tip

We strongly recommend performing your own hyperparameter search to achieve the best performance in your environment using the above provided search scripts.

📊 Datasets and Splits

We use the open-source pip package graph_datasets for unified data loading:

$ python -m pip install graph_datasets

Example usage:

from graph_datasets import load_data
from configs import DataConf
from utils import read_configs

DATA_INFO = DataConf(**read_configs("data"))

data = load_data(
    dataset_name='squirrel',
    source='critical',
    directory=DATA_INFO.DATA_DIR,
    row_normalize=True,
    rm_self_loop=False,
    add_self_loop=True,
    verbosity=1,
    return_type="pyg",
)

To minimize variance from inconsistent split policies across datasets, we use a unified 10× random split scheme with a 48%/32%/20% train/validation/test ratio.

• For medium-size datasets, the splits are stored in ./data/random_splits/fixed_splits/.
• For large datasets, OGB-arxiv and OGB_products are using the public splits with pokec using the splits from this work.
• All splits can be loaded via:

from utils import get_splits

# `repeat` is the number of our/public splits
# `i` is the index of the selected split
train_mask, val_mask, test_mask = get_splits(
    data=data,
    name=data.name,
    n_nodes=data.num_nodes,
    i=1,
    repeat=10,
    TRAIN_RATIO=48,
    VALID_RATIO=32,
    DATA=DATA_INFO,
    public=False,
)

🧩 Baselines

The code for all 30 baselines is in ./benchmark/baselines:

• If a baseline has its own folder, a search.py script is included for hyperparameter tuning with optuna. See the README.md in the folder for details.
• If a baseline does not have its own folder, it can be run with a script like ./baselines.py, which can conveniently derive the corresponding search.py script.
• All search spaces used in the experiments are documented in ./configs/search_grid.py.

📝 Empirical Analysis

• The code for the empirical analysis is documented in ./scripts/dml.sh.
• Run the analysis and draw the analysis plots via:

$ bash scripts/dml.sh squirrel critical False 10
$ python -u -m scripts.merge

📚 Citation

If you find this work useful, please cite our paper:

@inproceedings{ignn,
  title={Making Classic {GNN}s Strong Baselines Across Varying Homophily: A Smoothness{\textendash}Generalization Perspective},
  author={Ming Gu and Zhuonan Zheng and Sheng Zhou and Meihan Liu and Jiawei Chen and Qiaoyu Tan and Liangcheng Li and Jiajun Bu},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=IAGbhDARZd}
}