This repository contains the implementation of MEG (IJCNN 2021).

Usage

We assume miniconda (or anaconda) to be installed.

Install dependencies

Run the following commands:

source setup/install.sh [cpu | cu92 | cu101 | cu102]
conda activate meg

Train DGN

Train the DGN to be explained by running:

python train_dgn.py [tox21 | esol] <experiment_name>

Generate counterfactuals

To generate counterfactual explanations for a specific sample, run:

python train_meg.py [tox21 | esol] <experiment_name> --sample <INTEGER>

Results will be saved at runs/<dataset_name>/<experiment_name>/meg_output.

Bibtex

@inproceedings{numeroso2021,
      author={Numeroso, Danilo and Bacciu, Davide},
      booktitle={2021 International Joint Conference on Neural Networks (IJCNN)}, 
      title={MEG: Generating Molecular Counterfactual Explanations for Deep Graph Networks}, 
      year={2021},
      volume={},
      number={},
      pages={1-8},
      doi={10.1109/IJCNN52387.2021.9534266}
}