LEGIT: Local Explainer for deep Graph networks by Input perturbaTion

This repository contains the official implementation of LEGIT (TNNLS).

Install dependencies

    conda env create -f envs/cpu.yml

or

    conda env create -f envs/gpu.yml

Train the predictor

Now you can train the graph network (with default hyperparameters) to be explained by running:

python dgn.py train [tox21 | esol | cycliq] <experiment_name>

If you wish to tune the hyperparameters, refer to the model-selection file.

Executing this command will populate the directory runs, which is structured as follows:

runs/
└── <dataset_name>
    └── <experiment_name>
        ├── best_performance.json
        ├── ckpt
        │   └── <model>.pth
        ├── hyperparams.json
        ├── gen_output
        │   └── <sample>
        │       └── data.json
        │
        ├── plots
        │   └── events.out.tfevents.*
        └── splits
            ├── test.pth
            ├── train.pth
            └── val.pth

Get graph embeddings

If you wish to replicate results of (some of) the baselines, you'll be required to have the node embeddings of the training and test graphs, especially for CoGE. You can get them by running:

python dgn.py embed [tox21 | esol | cycliq] <experiment_name>

Embeddings will be saved to runs/<dataset_name>/<experiment_name>/embeddings/[train | test].

Train the generator

After the DGN has been trained, run:

python train_gen.py [tox21 | esol | cycliq] <experiment_name>

The generator will start searching for the best neighbours and will dump them to runs/<dataset_name>/<experiment_name>/gen_output/<sample>/data.json.

Generate explanations

Once we have the neighbours for each test sample you would like to explain, run:

python explain.py contrast | linear  <...parameters>

contrast only work for CYCLIQ, whereas linear requires chemistry tasks (i.e, TOX21, ESOL).

To evaluate explanation accuracy for CYCLIQ, we use the evaluation script from CoGE:

python evaluate.py <dataset_path> <explain_path>

For more information on the available explainers, run python explain.py --help or python explain.py <method> --help.