GenRe : Generative Models for Robust Algorithmic Recourse

1. setup environment using requirements.txt. Install PyTorch
2. setup datasets

chmod +x ./scripts/setup_datasets.sh
./scripts/setup_datasets.sh

3. train true classifiers

python scripts/train_rf.py [options]
# example
python scripts/train_rf.py --dataset heloc

4. train ann classifiers

python scripts/train_ann.py [options]
# example
python scripts/train_ann.py --dataset heloc --lsrc rf --epochs 100 --batch 64 --lr 0.001 --device 0 --hidd 10 10 10

5. train pairmodel

python scripts/train_bpm.py [options]

6. Checkout genre_sampler.ipynb for running the code and evaluation.

To see available options, use --help

Baselines

To run baselines, install CARLA from https://github.com/MartinPawel/ProbabilisticallyRobustRecourse/ and activate the environment[don't install GPU version of pytorch]

• Diverse Counterfactual Explanations (DiCE): Paper
• Growing Sphere (GS): Paper
• Wachter: Paper
• ROAR: Paper
• PROBE: Paper
• CCHVAE: Paper
• REVISE: Paper
• CRUDS: Paper

Datasets

• Adult: Source
• COMPAS: Source
• HELOC: Source

Citation

If you use this work, please cite:

@inproceedings{
garg2025sample,
title={From Search to Sampling: Generative Models for Robust Algorithmic Recourse},
author={Garg, Prateek and Nagalapatti, Lokesh and Sarawagi, Sunita},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=NtwFghsJne}
}

Contact

Prateek Garg (prateekg@iitb.ac.in)