GNN-Substructure-Counting

The official implementation of ''Can Graph Neural Networks Count Substructures?'' at NeurIPS 2020.

Authors: Zhengdao Chen (NYU), Lei Chen (NYU), Soledad Villar (JHU), Joan Bruna (NYU).

Dependencies

Core packages:

pytorch 1.5.0
dgl 0.4.3
torch-sparse 0.6.5
torch-scatter 2.0.4
ogb 1.1.1
scipy 1.2.1

Note that

1) DGL released a massive update of APIs in 0.5, due to which the implementation is in need of modification accordingly if you would like to install a more recent version. We refer to an official API guidebook for 0.4.x [1] and an official blog revealing what's new in 0.5 release [2].

2) Following the official instructions [3] to install torch-sparse according to versions of your pytorch and cuda. Note that torch-scatter is necessary althougth we do not explicitly import it in the implementation.

3) Other packages with higher verisions should be compatible. If you are with any questions, please do not hesitate to email us via lc3909@nyu.edu.

[1] https://docs.dgl.ai/en/0.4.x/api/python/graph.html
[2] https://www.dgl.ai/release/2020/08/26/release.html
[3] https://github.com/rusty1s/pytorch_sparse

For citation

@article{chen2020can,
  title={Can graph neural networks count substructures?},
  author={Chen, Zhengdao and Chen, Lei and Villar, Soledad and Bruna, Joan},
  journal={arXiv preprint arXiv:2002.04025},
  year={2020}
}

or

@inproceedings{NEURIPS2020_75877cb7,
 author = {Chen, Zhengdao and Chen, Lei and Villar, Soledad and Bruna, Joan},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {H. Larochelle and M. Ranzato and R. Hadsell and M. F. Balcan and H. Lin},
 pages = {10383--10395},
 publisher = {Curran Associates, Inc.},
 title = {Can Graph Neural Networks Count Substructures?},
 url = {https://proceedings.neurips.cc/paper/2020/file/75877cb75154206c4e65e76b88a12712-Paper.pdf},
 volume = {33},
 year = {2020}
}