GNN Explainability Framework

Node Classification Tasks

Explainer	Paper
PageRank	The PageRank Citation Ranking: Bringing Order to the Web
Distance	Shortest Path Distance Approximation using Deep learning Techniques
SA	Explainability Techniques for Graph Convolutional Networks.
Grad-CAM	Explainability Methods for Graph Convolutional Neural Networks.
DeepLIFT	Learning Important Features Through Propagating Activation Differences
Integrated Gradients	Axiomatic Attribution for Deep Networks
GNNExplainer	GNNExplainer: Generating Explanations for Graph Neural Networks
SubgraphX	On Explainability of Graph Neural Networks via Subgraph Exploration
PGM-Explainer	PGM-Explainer: Probabilistic Graphical Model Explanations for Graph Neural Networks

Graph Classification Tasks

Explainer	Paper
ReFine	Towards Multi-Grained Explainability for Graph Neural Networks
SA	Explainability Techniques for Graph Convolutional Networks.
Grad-CAM	Explainability Methods for Graph Convolutional Neural Networks.
DeepLIFT	Learning Important Features Through Propagating Activation Differences
Integrated Gradients	Axiomatic Attribution for Deep Networks
GNNExplainer	GNNExplainer: Generating Explanations for Graph Neural Networks
PGExplainer	Parameterized Explainer for Graph Neural Network
PGM-Explainer	PGM-Explainer: Probabilistic Graphical Model Explanations for Graph Neural Networks
Screener	Causal Screening to Interpret Graph Neural Networks
CXPlain	Cxplain: Causal Explanations for Model Interpretation under Uncertainty

Installation

Requirements

• CPU or NVIDIA GPU, Linux, Python 3.7
• PyTorch >= 1.5.0, other packages

1. Pytorch Geometric. Official Download.

# We use TORCH version 1.6.0
CUDA=cu111
TORCH=1.9.1
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html
pip install torch-geometric==2.0.3

2. Other packages

pip install tqdm matplotlib argparse json jupyterlab notebook pgmpy captum
# For visualization (optional)
pip install tensorboardx

Datasets

1. The processed raw data for datasets syn1, syn2, syn3, syn4, syn5, syn6 is available in the data/syn folder.
2. The processed raw data for datasets cora, citeseer, pubmed, cornell, texas, wisconsin, chameleon, squirrel, actor will be automatically downloaded when training models.

Trained GNN models

We provide the trained GNNs in model/ for reproducing the results in our paper.

Python code map

.
├── dataset
│   ├── __init__.py
│   ├── __pycache__
│   │   ├── data_utils.cpython-38.pyc
│   │   ├── gen_planetoids.cpython-38.pyc
│   │   └── gen_syn.cpython-38.pyc
│   ├── data_utils.py
│   ├── gen_mutag.py
│   ├── gen_real.py
│   ├── gen_syn.py
│   ├── mutag_utils.py
│   └── syn_utils
│       ├── featgen.py
│       ├── gengraph.py
│       ├── gengroundtruth.py
│       └── synthetic_structsim.py
├── evaluate
│   ├── __init__.py
│   ├── __pycache__
│   │   ├── accuracy.cpython-38.pyc
│   │   └── fidelity.cpython-38.pyc
│   ├── accuracy.py
│   ├── fidelity.py
│   └── mask_utils.py
├── explainer
│   ├── __init__.py
│   ├── genmask.py
│   ├── gnnexplainer.py
│   ├── graph_explainer.py
│   ├── node_explainer.py
│   ├── pgmexplainer.py
│   ├── shapley.py
│   └── subgraphx.py
├── gnn
│   ├── __init__.py
│   ├── eval.py
│   ├── model.py
│   └── train.py
├── main.py
└── utils
    ├── __init__.py
    ├── gen_utils.py
    ├── graph_utils.py
    ├── io_utils.py
    ├── math_utils.py
    ├── parser_utils.py
    └── plot_utils.py

Demo

Node Classification

python3 code/main.py --dataset [dataset-name] --explain_graph False --explainer_name [explainer_name]

• dataset-name:
  • synthetic: syn1, syn2, syn3, syn4, syn5, syn6
  • real-world: cora, pubmed, citeseer, facebook, chameleon, squirrel, texas, wisconsin, cornell, actor
• explainer_name: random, pagerank, distance, sa, ig, gradcam, occlusion, basic_gnnexplainer, gnnexplainer, subgraphx, pgmexplainer

Graph Classification

python3 code/main.py --dataset [dataset-name] --explain_graph True --explainer_name [explainer_name]

• dataset-name: mutag
• explainer_name: random, sa, ig, gnnexplainer (to complete)

Mask transformation

To compare the methods, we adopt separately three strategies to cut off the masks:

1. Sparsity

2. Threshold

3. Topk

Baseline explainers

gnn_explainer = GNNExplainer(device, gnn_path)
gnn_explainer.explain_graph(test_dataset[0],
                           epochs=100, lr=1e-2)

screener = Screener(device, gnn_path)
screener.explain_graph(test_dataset[0])

Visualization

Visualisation of the GNN model training Visualisation of the explanations

Jupyter Notebook

The default visualizations are provided in notebook/GNN-Explainer-Viz.ipynb.

Note: For an interactive version, you must enable ipywidgets

jupyter nbextension enable --py widgetsnbextension

Tuningthe mask sparsity/threshold/top-k values. You can now play around with the mask threshold in the GNN-Explainer-Viz-interactive.ipynb.

TODO: Explain outputs + visualizations + baselines

Included experiments

Name	EXPERIMENT_NAME	Description
Synthetic #1	syn1	Random BA graph with House attachments.
Synthetic #2	syn2	Random BA graph with community features.
Synthetic #3	syn3	Random BA graph with grid attachments.
Synthetic #4	syn4	Random Tree with cycle attachments.
Synthetic #5	syn5	Random Tree with grid attachments.
Synthetic #6	syn6	Random BA graph with bottle attachments.
MUTAG	mutag	Mutagenecity Predicting the mutagenicity of molecules (source).
Cora	cora	Citation network (source).
Pubmed	pubmed	PubMed network (source).
Citeseer	citeseer	Citeseer network (source).
FacebookPagePage	facebook
Chameleon	chameleon
Squirrel	squirrel
Texas	texas
Wisconsin	wisconsin
Cornell	cornell
Actor	actor

Using the explainer on other models

A graph convolutional model is provided. This repo is still being actively developed to support other GNN models in the future.

Changelog

See CHANGELOG.md

Citation

Please cite our paper if you find the repository useful.

@inproceedings{}