github.com/balcilar/Spectral-Designed-Graph-Convolutions

Spectral Designed Graph Convolutions

Codes of "Bridging the Gap Between Spectral and Spatial Domains in Graph Neural Networks" paper.

Requirements

These libraries versions are not stricly needed. But these are the configurations in our test machine.

• Python==3.6.5
• tensorflow-gpu==1.15.0
• numpy==1.17.4
• networkx==2.4
• scipy==1.3.1
• matplotlib==3.1.2
• pickle==4.0

Usage

Run the scripts directly. All parameters are defined in corresponding script. In Pubmed and PPI dataset, since the eigen decomposition takes quite a long time because of the dimension of given graph, we write the eigenvectors into file in first run. For later run, the code directly read already calculated eigenvectors from file.

Transductive Setting Problems

python cora_multirun.py
python citeseer_multirun.py
python pubmed_multirun.py

Inductive Setting Problems

python ppi_singlerun.py
python protein_nodelabel.py
python enzymes_nodelabel.py
python enzymes_allfeats.py

Results

Citation

Please cite this preprint paper if you want to use it in your work,

@article{balcilar2020bridging,
  title={Bridging the Gap Between Spectral and Spatial Domains in Graph Neural Networks},
  author={Balcilar,Muhammet  and Renton, Guillaume and H\'eroux,Pierre and Ga\"uz\`ere,Benoit and Adam, S\'ebastien and Honeine,Paul},
  journal={arXiv preprint arXiv:2003.11702},
  year={2020},
      eprint={2003.11702},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

And/or this published paper in ICLR2021 which covers the theoretical part.

@inproceedings{
balcilar2021analyzing,
title={Analyzing the Expressive Power of Graph Neural Networks in a Spectral Perspective},
author={Muhammet Balcilar and Guillaume Renton and Pierre H{\'e}roux and Benoit Ga{\"u}z{\`e}re and S{\'e}bastien Adam and Paul Honeine},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=-qh0M9XWxnv}
}

 

License

MIT License