GNNShap: Scalable and Accurate GNN Explanation using Shapley Values

This repository contains the source code of GNNShap: Scalable and Accurate GNN Explanation using Shapley Values paper accepted in The Web Conference 2024.

Setup

Our implementation is based on PyTorch and PYG. Also, Our Shapley sampling strategy is implemented in Cuda. Therefore, GNNShap requires a GPU with Cuda support.

First install PyTorch with GPU support from here and make sure PyTorch is using GPU.

The rest of the required packages and versions are provided in the requirements.txt file.

You can install the requirements by running:

pip install -r requirements.txt

Dataset Configs

Dataset and dataset-specific model configurations are in the dataset/configs.py file.

Model training

We provided pretrained models in the pretrained folder.

To train Cora, CiteSeer, PubMed, Coauthor-CS, Coauthor-Physics, and Facebook datasets:

python train.py --dataset Cora

Reddit and ogbn-products require NeighborLoader for training. To train them:

python train_large.py --dataset Reddit

Experiments

We provided scripts for baselines and GNNShap experiments. Scripts will save explanation results to the results folder. Note that scripts repeat each experiment five times. This can be changed in the scripts.

For baselines, you can use the following script. For individual baseline, you can refer to the script file content.

./run_baseline_experiments.sh

For GNNShap experiments, you can use the following script:

./run_gnnshap_experiments.sh

• We ran experiments on a GPU with 24GB of memory. You may need to adjust batch_size and num_samples parameters if you have less GPU memory.
• The first run might take some time: it needs to compile the Cuda code.

For individual dataset experiments, an example is provided below:

python run_gnnshap.py --dataset Cora --num_samples 25000 --sampler GNNShapSampler 
--batch_size 1024 --repeat 1

The results will be saved to the results folder. The default result folder can be changed in dataset/configs.py

Evaluation

We used the BaselinesEvaluation.ipynb notebook under the examples folder for explanation times and fidelity results.

Visualization

We provided explanation visualization examples in the Visualization.ipynb notebook under examples.

Custom Model & Data Explanations

We provided an example in the CustomModelData.ipynb notebook under examples.

Citation

Please cite our work if you find it useful.

Selahattin Akkas and Ariful Azad. 2024. GNNShap: Scalable and Accurate GNN Explanation using Shapley Values.
In Proceedings of the ACM Web Conference 2024 (WWW ’24), May 13–17, 2024, Singapore, Singapore.