D4Explainer: In-distribution Explanations of Graph Neural Network via Discrete Denoising Diffusion [NeurIPS 2023]

This is the Pytorch implementation of " D4Explainer: In-distribution Explanations of Graph Neural Network via Discrete Denoising Diffusion"

Requirements

• torch==1.10.1
• torch-geometric==2.0.4
• numpy==1.24.2
• pandas==1.5.3
• networkx==3.0

Refer to requirements.txt for more details.

Dataset

Download the datasets from here to data/

Datasets Included:

• Node classification: BA_shapes; Tree_Cycle; Tree_Grids; cornell
• Graph classification: mutag; ba3; bbbp; NCI1

Train Base GNNs

cd gnns
python ba3motif_gnn.py
python bbbp_gnn.py
python mutag_gnn.py
python nci1_gnn.py
python synthetic_gnn.py --data_name Tree_Cycle
python synthetic_gnn.py --data_name BA_shapes
python tree_grids_gnn.py
python web_gnn.py

Train and Evaluate D4Explainer

For example, to train D4Explainer on Mutag, run:

python main.py --dataset mutag

Evaluation of Other Properties

• In-distribution: python -m evaluation.ood_evaluation
• Robustness: python -m evaluation.robustness