L2C

This repo contains codes for reproducing the experiments in the paper Learning to Counter: Feature-based Learning for Diverse and Privacy-Preserving Counterfactual Explanations accepted at KDD 2023.

Dependencies

torch, sklearn and pandas==1.3.5

Data

The folder data/ contains train/test splits for each dataset : German Credit (german), Graduate Admission (admission), Student Performance (student) and Small Business Administration (sba).

Our experiment setup is described in data_generator.py and data_loader.py.

Black-box Models

Each dataset has its own directory in data/ and model/

mkdir model
for name in german admission student sba
do
  mkdir model/$name/
  mkdir data/$name/
done

Refer to blackbox.py for how to train, save and load pre-trained blackbox models. Please ensure the models are stored in the folders corresponding to the dataset names.

Running L2C

To train the generative model, run the following command

python main.py <dataset-name> train

Please specify the correct dataset name given in the parentheses above. If there is no pre-trained black-box model, it will be trained before L2C model is run. Refer to data_loader.py and blackbox.py for details.

Baselines

Reproduction for FastAR can be found in FastAR/. Refer to FastAR repo for more explanations.

First, you need to transfer our pre-processed data and pre-trained black-box in FastAR/fastar/datasets and FastAR/fastar/models by running

python compare.py fastar None

Next, go the this directory. Install dependencies and set up the sub-directories

cd FastAR
bash install_requirements.sh
cd fastar
for name in german admission student sba
do
  mkdir output/samples/$name/
  mkdir output/trained_models/$name/
done

You can find the base scripts run_<dataset-name>.sh for training and evaluation. Since FastAR only supports single class, the variable $y is to specify the desired class. Before running the script, you need to adjust line 61 and line 76 in gym-midline/gym_midline/envs/my_<dataset-name>.py accordingly. The current setup assumes label 0 is undesirable class and 1 is the desired outcome. The outputs will be written in FastAR/fastar/output/.

bash run_german.sh train
bash run_german.sh eval

In baselines/, we provide our customized codes for other baselines, gratefully borrowed from DiCE repo, MCCE repo, COPA repo and Certifai repo. Feasible-VAE is supported in DiCE library. To run and evaluate the DiCE, F-VAE, COPA, MCCE and Certifai,

python compare.py <method-name> <dataset-name>

CRUDS can be implemented off-the-shelf from CARLA repo.

Citation

If you use the codes or datasets in this repository, please cite our paper.

@inproceedings{vo2023feature,
  title={Feature-based Learning for Diverse and Privacy-Preserving Counterfactual Explanations},
  author={Vo, Vy and Le, Trung and Nguyen, Van and Zhao, He and Bonilla, Edwin V and Haffari, Gholamreza and Phung, Dinh},
  booktitle={Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={2211--2222},
  year={2023}
}