FlowX: Towards Explainable Graph Neural Networks via Message Flows

Dependencies

See environment.yml and requirements.txt. The code is also tested on PyTorch 1.10.1, PyG 2.0.4.

Run FlowX

python -m benchmark.kernel.pipeline.py --task explain --model_name GCN_3l --dataset_name bbbp --target_idx 0 --explainer FlowX_plus --sparsity 0.7 --force_recalculate

License

This project is licensed under the terms of the MIT license.

Citing FlowX

If you find FlowX useful in your research, please consider citing:

@article{gui2022flowx,
  title={Flowx: Towards explainable graph neural networks via message flows},
  author={Gui, Shurui and Yuan, Hao and Wang, Jie and Lao, Qicheng and Li, Kang and Ji, Shuiwang},
  journal={arXiv preprint arXiv:2206.12987},
  year={2022}
}