networks-learning/counterfactual-explanations-mdp

Code for "Counterfactual Explanations in Sequential Decision Making Under Uncertainty", NeurIPS 2021

Counterfactual Explanations in Sequential Decision Making Under Uncertainty

This repository contains the code used in the paper Counterfactual Explanations in Sequential Decision Making Under Uncertainty, published at NeurIPS 2021.

Dependencies

All the experiments were performed using Python 3.9. In order to create a virtual environment and install the project dependencies you can run the following commands:

python3 -m venv env
source env/bin/activate
pip install -r requirements.txt

Code organization

The directory src contains the source code for the various experiments. The following table contains a short description for each python file:

src/	Description
synth_mdp.py	Defines a Synth_MDP object which handles the generation of synthetic decision making realizations and their counterfactual explanations.
synth_compute_cf_mdps.py	Generates synthetic decision making realizations, pre-computes the counterfactual transition probabilties for each one of them and saves them as .pkl files.
synth_experiment.py	Generates optimal counterfactual explanations for each synthetic decision making realization, for various values of the parameter k.
therapy_mdp.py	Defines a Therapy_MDP object which handles the input of the cognitive behavioral therapy data and the generation of counterfactual explanations for each patient.
therapy_compute_cf_mdps.py	Pre-computes the counterfactual transition probabilties for all patients' realizations and saves them as .pkl files.
therapy_experiment.py	Generates optimal counterfactual explanations for each patient's realization, for various values of the parameter k.
therapy_evaluation.py	Generates counterfactual explanations for each patient's realization, using our method and various baselines.

The directory scripts contains bash scripts that use the aforementioned code and pass parameter values required for the various experiments.

The directory notebooks contains jupyter notebooks producing the figures appearing in the paper. Each notebook has script execution prerequisites specified therein.

The directory figures is used for saving the figures produced by the notebooks.

The directory outputs is used for saving the json outputs produced by the scripts. The sub-directory cf_mdps is used for saving intermediate .pkl files containing the counterfactual transition probabilities for each single realization.

Citation

If you use parts of the code in this repository for your own research, please consider citing:

@inproceedings{tsirtsis2021counterfactual,
    title={Counterfactual Explanations in Sequential Decision Making Under Uncertainty},
    author={Tsirtsis, Stratis and De, Abir and Gomez-Rodriguez, Manuel},
    booktitle={Advances in Neural Information Processing Systems},
    year={2021}
}