Decouple Graph Neural Networks: Train Multiple Simple GNNs Simultaneously Instead of One

This repository is our implementation of

Hongyuan Zhang, Yanan Zhu, and Xuelong Li, "Decouple Graph Neural Networks: Train Multiple Simple GNNs Simultaneously Instead of One," IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), DOI: 10.1109/TPAMI.2024.3392782, 2024.(arXiv)(IEEE)

SGNN attempts to further reduce the training complexity of each iteration from $\mathcal{O}(n^2) / \mathcal{O}(|\mathcal E|)$ (vanilla GNNs without acceleration tricks, e.g., AdaGAE) and $\mathcal O(n)$ (e.g., AnchorGAE) to $\mathcal O(m)$.

Compared with other fast GNNs, SGNN can

• (Exact) compute representations exactly (without sampling);
• (Non-linear) use up to $L$ non-linear activations ($L$ is the number of layers);
• (Fast) be trained with the real stochastic (mini-batch based) optimization algorithms.

The comparison is summarized in the following table.

If you have issues, please email:

hyzhang98@gmail.com

Requirements

• pytorch 1.10.0
• scipy 1.3.1
• scikit-learn 0.21.3
• numpy 1.16.5

How to run SGNN

Please ensure the data is rightly loaded

python run.py
python run_classfication.py

Citation

@article{SGNN,
  author={Zhang, Hongyuan and Zhu, Yanan and Li, Xuelong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Decouple Graph Neural Networks: Train Multiple Simple GNNs Simultaneously Instead of One}, 
  year={2024},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2024.3392782}
}