ShichangZh / GStarX

[NeurIPS 2022] Explaining Graph Neural Networks with Structure-Aware Cooperative Games (GStarX)

Graph Structure-aware Explanation (GStarX)

Code for Explaining Graph Neural Networks with Structure-Aware Cooperative Games by Shichang Zhang, Yozen Liu, Neil Shah, and Yizhou Sun.

Visualizations

Explanation on the graph-sst2 dataset

GStarX concisely captures the important words for sentiment classification without including extraneous ones for either the positive sentence or the negative sentence.

Explanation on the mutag dataset

GStarX generates explanations with much better Fidelity/Inv-Fidelity/H-Fidelity than other methods and even better than the “ground truth” specified by the domain knowledge.

Getting Started

Requirements

• Please follow the links below to install PyTorch and PyG with proper CUDA versions

  • PyTorch https://pytorch.org/
  • PyG https://github.com/pyg-team/pytorch_geometric

• Then install packages by running the line below

pip install -r requirements.txt

• Our code has been tested with
  • Python = 3.8.10
  • PyTorch = 1.10.0
  • PyG = 1.7.2

Datasets

To run experiments for datasets used in the paper, please download them from this link and put them under datasets/ (all dataset folder names should be in lowercase, some of them need to be renamed. Please be consistent with dataset.py).

You may also add your favourite datasets by modifying the get_dataset function in dataset.py.

Models

We borrow the GCN, GIN, and GAT implementation and the model training script train_gnns.py from the DIG library. Pre-trained checkpoints of these models are stored in checkpoints/

Usage

• Run the GStarX to explain trained GNN models
  • A simple example is shown below, and more examples are in script.sh

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

• Run a baslineline method to explain trained GNN models
  • Run baseline scripts like the example below by replacing method with gnn_explainer, pgexplainer, subgraphx, graphsvx, or orphicx.

python baselines/run_${method}.py models='gcn' datasets='bace'

• Train new GNNs to explain
  • If not GCN, GIN, or GAT, add the model architecture to getNets.py, and add its configuration to config
  • Run train_gnns.py as the example below

python train_gnns.py datasets=ba_2motifs models=gat

Results

GStarX vs. other methods for graph prediction explanation in terms of the H-Fidelity metric. GStarX outperforms others on most datasets, on average, and on different GNN architectures. (Please see Table 1 and Table 2 in the paper).

Datasets	GNNExplainer	PGExplainer	SubgraphX	GraphSVX	OrphicX	GStarX
BA_2Motifs (GCN)	0.4841	0.4879	0.6050	0.5017	0.5087	0.5824
BACE (GCN)	0.5016	0.5127	0.5519	0.5067	0.4960	0.5934
BBBP (GCN)	0.4735	0.4750	0.5610	0.5345	0.4893	0.5227
GraphSST2 (GCN)	0.4845	0.5196	0.5487	0.5053	0.4924	0.5519
MUTAG (GCN)	0.4745	0.4714	0.5253	0.5211	0.4925	0.6171
Twitter (GCN)	0.4838	0.4938	0.5494	0.4989	0.4944	0.5716
Average (GCN)	0.4837	0.4934	0.5569	0.5114	0.4952	0.5732

Datasets	GNNExplainer	PGExplainer	SubgraphX	GraphSVX	OrphicX	GStarX
GraphSST2 (GAT)	0.4951	0.4918	0.5484	0.5132	0.4997	0.5542
MUTAG (GIN)	0.5042	0.4993	0.5264	0.5592	0.5152	0.6064

Cite

Please cite our paper if you find this code is useful. Thank you.

@misc{zhang2022explaining,
      title={Explaining Graph-level Predictions with Communication Structure-Aware Cooperative Games}, 
      author={Shichang Zhang and Neil Shah and Yozen Liu and Yizhou Sun},
      year={2022},
      eprint={2201.12380},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Reference

Our code heavily relies on the DIG library. We encourage readers to check it out:

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

Contact Us

Please open an issue or contact [e-mail] if you have any questions.