🔥SAME🔥

This repo contains PyTorch implementation of the paper: SAME: Uncovering GNN Black Box with Structure-aware Shapley-based Multipiece Explanations in 2023 NeurIPS.

Getting Started

Requirements

The recommended Python version is 3.8 or 3.9. (SAME is tested under Python 3.8, Pytorch 1.12.1 and PyG 2.1.0)

conda create -n same python==3.8

Then, please follow the Pytorch Guidelines to install Pytorch and INSTALL PYG to install PyG. After that, install other packages using:

pip install -r requirements.txt

Datasets

For the datasets used in our paper, please download them manually and move them into the folder datasets/. The file name is case sensitive and please make sure that all of the file names are in lowercase. For more information about the datasets, please refer to dataset.py.

Here we provide the link to download them. (SAME_datasets contains both raw and processed data. If you want raw data only, please kindly click SAME_raw_datasets)
SAME_datasets FILE_ID=1VXfb6hnKZ3nBsKMJPqvjI4Nz0ikk8skT
SAME_raw_datasets FILE_ID=1KqSHf6Xz_PXKNA6KwXucye1H18hKIMJ6

[Easy for Terminal] You can download the dataset using the following bash. Remember to replace FILE_ID with that shown above.

# install gdown to down file from google drive
pip install gdown
# download file using FILE_ID
gdown https://drive.google.com/uc?id=FILE_ID

Models

Please use the following bash script to train the GNN models. Our code contains the implementation of GCN, GAT and GIN. The trained checkpoints of these models will be saved in checkpoints/.

python train_gnns.py models='gcn' datasets='mutag'

Usage

• Train GNNs before explaining.
  • The models argument can be chosen from GCN, GAT or GIN. And you can create your own GNN model in getNets.py.
  • For other datasets and GNNs, please create the corresponding files in config.

python train_gnns.py datasets='mutag' models='gin'

• One-Click run: use the following script to run the experiments thoroughly. The script will automatically run the experiments in the sparsity $[0.5,0.6,0.7,0.8]$.

bash main.sh

• Otherwise, you can run one explainer to explain a trained GNN model
  • You can run other baseline methods by setting explainer as gnn_explainer, pgexplainer, gnn_lrp, grad_cam, subgraphX, gstarx, orphicx, and graphsvx

explainer=same
dataset_name=bbbp
sparsity=0.8
model_name=gcn
bash run_one_model.sh $explainer $dataset_name $sparsity $model_name

Acknowledgement

Some baseline methods of our code are adapted from the following repositories.

https://github.com/ShichangZh/GStarX

https://github.com/divelab/DIG/tree/main/dig/xgraph

Contact Us

If you have any questions, please feel free contact us: [email protected] and rihanhuang.work@gmail or open an issue.