Counterfactual Explanation_vs_Adversarial Examples

This is the code repository for paper: "Reproducibility and Comparative Analysis of GNN Counterfactual Explanations with Adversarial Examples"

Table of Contents

• Requirements
• CFF
• CF_GNNExplainer
• CLEAR
• GCFExplainer
• RCExplainer
• Adversarials

Requirements

• Python 3.10

• pytorch 1.13.0

• cuda 11.7

• dgl-cuda10.2 (Edit graphconv.py)

• torch-geometric 2.2

  Detailed requirements are available in requirements.txt in root folder and discussed more in each method directory.

CFF

1. To run the experiments, dgl-cuda library is required. https://docs.dgl.ai/install/index.html.
2. After installing dgl-cuda library, Please replace the graphconv.py file in the dgl library with the file (with the same name) provided in the CE/gnn_cff folder. This is for the relaxation purpose as described in the paper.
3. The training and explaining are independent phases. For you convenience, we provide pre-trained GNN model params. Under the CE/gnn_cff folder, run:

   unzip log.zip
   

4. To set the python path, under the project root folder, run the following command:

   export PYTHONPATH=$PYTHONPATH:"$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )/CE/gnn_cff"
   

5. As an example, to generate explanations for node classification task (Tree-Cycles dataset as example), run:

   python CE/gnn_cff/scripts/exp_node_tree_cycles.py
   

6. For generating explanations for graph classification task (Mutag0 dataset as example), run:

   python CE/gnn_cff/scripts/exp_graph.py
   

7. The codes for training GNNs are also provided. For instance, for training GNN for node classification, run:

   python CE/gnn_cff/scripts/train_node_classification.py
   

   for graph classification, run:

   python CE/gnn_cff/scripts/train_graph_classification.py
   

CF-GNNExplainer

Training original models

To train the original GNN models for the BA-shapes dataset in the paper, cd into CE/cf-gnnexplainer/src and run this command:

python train.py --dataset=syn1

📋 For the Tree-Cycles dataset, the dataset argument should be "syn4". For the Tree-Grid dataset, it should be "syn5". All hyperparameter settings are listed in the defaults, and all models have the same hyperparameters.

Training CF-GNNExplainer

To train CF-GNNExplainer for each dataset, run the following commands:

python main_explain.py --dataset=syn1 --lr=0.01 --beta=0.5 --n_momentum=0.9 --optimizer=SGD
python main_explain.py --dataset=syn4 --lr=0.1 --beta=0.5 --optimizer=SGD
python main_explain.py --dataset=syn5 --lr=0.1 --beta=0.5 --optimizer=SGD

📋 This will create another folder in the main directory called 'results', where the results files will be stored.

Evaluation

To evaluate the CF examples, run the following command:

python evaluate.py --path=../results/<NAME OF RESULTS FILE>

📋 This will print out the values for each metric.

Pre-trained Models

The pretrained models are available in the models folder

Results

Our model achieves the following performance:

Model name	Dataset	Fidelity	Size	Sparsity	Accuracy
CF-GNNExplainer	Tree-Cycles	0.21	2.09	0.90	0.94
CF-GNNExplainer	Tree-Grid	0.07	1.47	0.94	0.96
CF-GNNExplainer	BA-Shapes	0.39	2.39	0.99	0.96

GCFExplainer

We have already provided gnn and neurosed base models. If you want to run this method using your own dataset, firstly, you have to train your own gnn and neurosed base models.

• For gnn base models, you can use our gnn.py module.
• For neurosed base models, please follow neurosed repository.

If neurosed model is hard to train, you will have to update our importance function to use your graph edit distance function.

Generating Counterfactual Candidates

To generate counterfactual candidates for AIDS dataset with the default hyperparameters, run this command:

python vrrw.py --dataset aids

The counterfactual candidates and meta information is saved under results/{dataset}/runs/. You can check other available training options with:

python vrrw.py --help

Generating Summary Counterfactuals

To generate counterfactual summary set for AIDS dataset from the candidates with the default hyperparameters, run this command:

python summary.py --dataset aids

The coverage and cost performance under different number of summary size will be printed on screen. You can check other available summary options with:

python summary.py --help

CLEAR

Dataset

Datasets can be found in link.

Run Experiment

Step 1: Training a graph prediction model

Train a graph prediction model (i.e., the model which needs explanation). The trained prediction models used in this paper can be directly loaded from ./model_save/.

If you want to train them from scratch, run the following command (here we use the dataset imdb_m as an example):

python train_pred.py  --dataset imdb_m --epochs 600 --lr 0.001

Or you can also use any other graph prediction models instead.

Step 2: Generating counterfactual explanations

python main.py --dataset imdb_m --experiment_type train

Here, when experiment_type is set to train or test, the model CLEAR will be trained or loaded from a saved file. When it is set to baseline, you can run the random perturbation based baselines (INSERT, REMOVE, RANDOM) by setting baseline_type.

RCExplainer

Datasets

Datasets are NOT required separately for inference or for training of the explanation methods. For ease, we have clubbed the datasets with the trained GNN pytorch models. Download trained models from this link. https://drive.google.com/file/d/14Tlv_beU8sGVk22AKsgseRF6dJEkXJAn/view?usp=sharing

Place 'ckpt' folder in 'RCExplainer/gcn_interpretation/gnn-model-explainer/'.

Datasets are ONLY required for training GNN model from scratch and are available for download from the following link if required: https://drive.google.com/file/d/1k3VjFPvHWM71Lfb9AXjHcC1MlJkMvoX5/view?usp=sharing For training GNN from scratch, please refer training GNN section. A copy of the datasets should be placed in RCExplainer/gcn_interpretation/datasets for this.

Pre-trained Models

Download all trained models from this link: https://drive.google.com/file/d/14Tlv_beU8sGVk22AKsgseRF6dJEkXJAn/view?usp=sharing and place 'ckpt' folder in RCExplainer/gcn_interpretation/gnn-model-explainer

GNN model + dataset is saved in RCExplainer/gcn_interpretation/gnn-model-explainer/ckpt/{dataset_name}/{dataset_name}_base.pth.tar

Explainer models are saved in RCExplainer/gcn_interpretation/gnn-model-explainer/ckpt/{dataset_name}/{method_name}/{method_name}.pth.tar

Explaining Trained GNN Models

This section covers how to evaluate already trained GNN models

Evaluating for fidelity-sparsity

The graph classification datasets are described below. We evaluate fidelity and sparsity on these datasets.

Dataset Name	Reference name	Ckpt dir	Explainer dir	Relevant class
BA-2Motifs	BA_2Motifs	ckpt/BA_Motifs_3gc	TBD	1
Mutagenicity	Mutagenicity	ckpt/orig_Mutagenicity	TBD	0
NCI1	NCI1	ckpt/NCI3_3gc	TBD	1

Below are commands for different datasets

1. Mutagenicity

python explainer_main.py --bmname Mutagenicity --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/Mutagenicity" --exp-path "./ckpt/Mutagenicity/RCExplainer/rcexplainer.pth.tar" --multigraph-class 0 --eval

2. BA_2Motifs

python explainer_main.py --bmname BA_2Motifs --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/BA_2Motifs" --exp-path "./ckpt/BA_2Motifs/RCExplainer/rcexplainer.pth.tar" --multigraph-class 1 --eval

3. NCI1

python explainer_main.py --bmname NCI1 --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/NCI1" --exp-path "./ckpt/NCI1/RCExplainer/rcexplainer.pth.tar" --multigraph-class 1 --eval

General method for running:

python explainer_main.py --bmname {Reference name} --num-gc-layers 3 --explainer-method rcexplainer --gpu --eval --graph-mode --ckptdir {Ckpt dir}  --exp-path  {Explainer dir} --multigraph-class {Relevant class}

Evaluating for noise robustness

For noise robustness results, simply add the --noise flag to the above commands. For example, for graph classification:

python explainer_main.py --bmname {Reference name} --num-gc-layers 3 --explainer-method rcexplainer --gpu --eval --graph-mode --ckptdir {Ckpt dir}  --exp-path  {Explainer dir} --multigraph-class {Relevant class} --noise

Example:

python explainer_main.py --bmname Mutagenicity --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/Mutagenicity" --exp-path "./ckpt/Mutagenicity/RCExplainer/rcexplainer.pth.tar" --multigraph-class 0 --eval

Evaluating node classification for accuracy and AUC

The node classification datasets are described below. We refer to RCExplainer/gcn_interpretation/gnn-model-explainer/ckpt as ckpt

Dataset Name	Reference name	Ckpt dir	Explainer dir	AUC	Node Accuracy
BA-Shapes	syn1	ckpt/syn1	TBD	0.998	0.973
BA-Community	syn2	ckpt/syn2	TBD	0.995	0.916
Tree-Grid	syn3	ckpt/syn3	TBD	0.993	0.993
Tree-Cycles	syn4	ckpt/syn4	TBD	0.995	0.974

Below are commands for node classification datasets

1. syn1 dataset (BA-Shapes)

python explainer_main.py --bmname syn1 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn1" --exp-path "./ckpt/syn1/RCExplainer/rcexplainer.pth.tar" --eval

2. syn2 dataset (BA-Community)

python explainer_main.py --bmname syn2 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn2" --exp-path "./ckpt/syn2/RCExplainer/rcexplainer.pth.tar" --eval

3. syn3 dataset (Tree-Grid)

python explainer_main.py --bmname syn3 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn3" --exp-path "./ckpt/syn3/RCExplainer/rcexplainer.pth.tar" --eval

4. syn4 dataset (Tree-Cycles)

python explainer_main.py --bmname syn4 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn4" --exp-path "./ckpt/syn4/RCExplainer/rcexplainer.pth.tar" --eval

General method for running:

python explainer_main.py --bmname {Dataset name} --num-gc-layers 3 --explainer-method rcexplainer --gpu --eval --ckptdir {Ckpt dir}  --exp-path  {Explainer dir}

Training Explainers

To train an explainer, you will have to pass in appropriate hyperparameters for each method.

For RCExplainer, this may include the learning rate --lr 0.001, the boundary and inverse boundary coefficient (which make up lambda) --boundary_c 3 --inverse_boundary_c 12, the size coefficient --size_c 0.01, the entropy coefficient --ent_c 10, and the boundary loss version --bloss-version "sigmoid" and name of folder where logs and trained models should be stored --prefix training_dir. The logs and models will be saved at RCExplainer/gcn_interpretation/gnn-model-explainer/ckpt/{dataset_name}/{prefix}/

Graph classification

Below are commands for training

1. Mutagenicity

python explainer_main.py --bmname Mutagenicity --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/Mutagenicity" --multigraph-class 0  --prefix "rcexp_mutag" --lr 0.001 --size_c 0.001 --ent_c 8.0 --boundary_c 3 --inverse_boundary_c 12 --bloss-version "sigmoid" --epochs 601

2. BA_2Motifs

python explainer_main.py --bmname BA_2Motifs --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/BA_2Motifs" --multigraph-class 1 --prefix "rcexp_ba2motifs" --lr 0.001 --size_c 0.001 --ent_c 8.0 --boundary_c 3 --inverse_boundary_c 12 --bloss-version "sigmoid" --epochs 601

3. NCI1

python explainer_main.py --bmname NCI1 --graph-mode --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/NCI1" --multigraph-class 1 --prefix "rcexp_nci1" --lr 0.001 --size_c 0.001 --ent_c 8.0 --boundary_c 3 --inverse_boundary_c 12 --bloss-version "sigmoid" --epochs 601

Node classification

Below are commands for training

Training explainer on syn2 (BA-Community)

python explainer_main.py --bmname syn2 --num-gc-layers 3 --explainer-method rcexplainer --gpu --lr 0.001 --boundary_c 10 --inverse_boundary_c 5 --size_c 0.09 --ent_c 10.0 --bloss-version "sigmoid" --topk 4 --prefix test-rc-syn2 --ckptdir "./ckpt/syn2_emb/"

1. syn1 dataset (BA-Shapes)

python explainer_main.py --bmname syn1 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn1 --lr 0.001 --boundary_c 12.0 --inverse_boundary_c 3.0 --size_c 0.09 --ent_c 10.0 --bloss-version "sigmoid" --prefix test-rc-syn1" --epochs 601

2. syn2 dataset (BA-Community)

python explainer_main.py --bmname syn2 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn2"  --lr 0.001 --boundary_c 12.0 --inverse_boundary_c 3.0 --size_c 0.09 --ent_c 10.0 --bloss-version "sigmoid" --prefix test-rc-syn2" --epochs 601

3. syn3 dataset (Tree-Grid)

python explainer_main.py --bmname syn3 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn3"  --lr 0.001 --boundary_c 12.0 --inverse_boundary_c 3.0 --size_c 0.09 --ent_c 10.0 --bloss-version "sigmoid" --prefix test-rc-syn3" --epochs 601

4. syn4 dataset (Tree-Cycles)

python explainer_main.py --bmname syn4 --num-gc-layers 3 --explainer-method rcexplainer --gpu --ckptdir "./ckpt/syn4" --lr 0.001 --boundary_c 12.0 --inverse_boundary_c 3.0 --size_c 0.09 --ent_c 10.0 --bloss-version "sigmoid" --prefix test-rc-syn4" --epochs 601

Other baselines

For training other baselines, replace --explainer-method rcexplainer in above commands with --explainer-method {baseline} where baseline can be picked from these methods gnnexplainer | pgexplainer | pgmexplainer | rcexp_noldb.

Some of the trained baselines and info can be downloaded from this url: https://drive.google.com/file/d/1t1i8kNjtGkqhI43ehGFvlxvbEan7HL80/view?usp=sharing For inference on these methods, simply change --exp-path in above inference commands to the provide explainer models. Inference of pgexplainer and rcexp_noldb baseline models is supported by rcexplainer, so keep the --explainer-method rcexplainer in inference commands.

Training GNN models from scratch

To train the GNN models from scratch, more setup is required:

Dependencies

It is necessary to install [pytorch geometric][pytorch-geometric.readthedocs.io/en/latest/notes/installation.html] matching your CUDA (10.1) and PyTorch version (1.8.0).

Please run the following to install this package:

--trusted-host and --no-cache-dir are required to circumvent network restrictions

Please replace TORCH and CUDA versions as appropriate.

conda install cudatoolkit-dev=10.1 -c pytorch
export TORCH=1.8.1
export CUDA=10.1
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html --trusted-host pytorch-geometric.com --no-cache-dir
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html --trusted-host pytorch-geometric.com --no-cache-dir
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html --trusted-host pytorch-geometric.com --no-cache-dir
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-${TORCH}+${CUDA}.html --trusted-host pytorch-geometric.com --no-cache-dir
pip install torch-geometric --trusted-host pytorch-geometric.com --no-cache-dir

Training

A model can be trained on a node classification dataset by:

python train.py --dataset syn1 --gpu

A model can be trained on a graph classification dataset by:

python train.py --bmname Mutagenicity --gpu --datadir {/path/to/benchmark/dataset}

For more details on training GNN models from scratch, check this GNNExplainer url https://github.com/RexYing/gnn-model-explainer