GMT: Graph Multilinear neT

This repo contains the sample code for reproducing the results of our ICML 2024 paper: How Interpretable Are Interpretable Graph Neural Networks?, which has also been presented as spotlight at ICLR MLGenX. 😆😆😆

Updates:

• Camera-ready version of the paper have been updated!
• Full code and instructions have been released!

Preparation

Environment Setup

We mainly use the following key libraries with the cuda version of 11.3:

torch==1.10.1+cu113
torch_cluster==1.6.0
torch_scatter==2.0.9
torch_sparse==0.6.12
torch_geometric==2.0.4

To setup the environment, one may use the following commands under the conda environments:

# Create your own conda environment, then...
conda install pytorch==1.10.1 torchvision==0.11.2 torchaudio==0.10.1 cudatoolkit=11.3 -c pytorch -c conda-forge
# Pytorch geometric
pip install torch_scatter==2.0.9 torch_sparse==0.6.12 torch_cluster==1.6.0 -f https://data.pyg.org/whl/torch-1.10.1+cu113.html
pip install torch_geometric==2.0.4
# Additional libraries
pip install -r requirements.txt

Datasets

To prepare the datasets of regular graphs, following the instructions in GSAT.

To prepare the datasets of geometric graphs, following the instructions in LRI.

Experiments on Regular Graphs

/GSAT contains the codes for running on regular graphs. The instructions to reproduce our results are given in scripts/gsat.sh.

Sample Commands

For GSAT

python run_gmt.py --dataset spmotif_0.5 --backbone GIN --cuda 0 -fs 1 -mt 0

For GMT-lin

python run_gmt.py --dataset spmotif_0.5 --backbone GIN --cuda 0 -fs 1 -mt 3 -ie 0.5

For GMT-sam

# train subgraph extractor
python run_gmt.py --dataset spmotif_0.5 --backbone GIN --cuda 0 -fs 1 -mt 5 -st 200 -ie 0.5 -sm 
# train subgraph classifier
python run_gmt.py --dataset spmotif_0.5 --backbone GIN --cuda 0 -fs 1 -mt 5550 -st 200 -ie 0.5 -fm -sr 0.8

Experiments on Geometric Graphs

/LRI contains the codes for running on geometric graphs. The instructions to reproduce our results are given in scripts/lri.sh.

Sample Commands

For LRI-Bern

python trainer.py -ba --cuda 0 --backbone egnn --dataset actstrack_2T --method lri_bern -mt 0

For GMT-lin

python trainer.py -ba --cuda 0 --backbone egnn --dataset actstrack_2T --method lri_bern -mt 0 -ie 0.1

For GMT-sam

# train subgraph extractor
python trainer.py -ba -smt 55 -ie 0.1 -fr 0.7 --cuda 0 --backbone egnn --dataset actstrack_2T --method lri_bern -mt 55 -ir 1
# train subgraph classifier
python trainer.py -ba -smt 55 -ie 0.1 -fr 0.7 --cuda 0 --backbone egnn --dataset actstrack_2T --method lri_bern -mt 5553

Misc

If you find our paper and repo useful, please cite our paper:

@inproceedings{chen2024gmt,
    title={How Interpretable Are Interpretable Graph Neural Networks?},
    author={Yongqiang Chen and Yatao Bian and Bo Han and James Cheng},
    booktitle={International Conference on Machine Learning},
    year={2024},
    url={https://openreview.net/forum?id=F3G2udCF3Q}
}

We would like to acknowledge the contribution of GSAT and LRI from Graph-COM to the base codes.