FAIR

Fair Adversarial Instance Re-weighting (Petrovic et al. https://arxiv.org/abs/2011.07495)

Overview

Here we provide the implementation of a Fair Adversarial Instance Re-weighting (FAIR) algotithm in PyTorch, along with a minimal execution examples (on the aif360 dataset). Besides four different variants of FAIR (FAIR-scalar, FAIR-betaREP, FAIR-betaSF and FAIR-Bernolli) we provided three different baselines:

Learning Unbiased Representations via Mutual Information backpropagation https://arxiv.org/abs/2003.06430
Conditional Learning of Fair Representations https://arxiv.org/abs/1910.07162
Fair Adversial Discriminative model http://mlg.eng.cam.ac.uk/adrian/AAAI2019_OneNetworkAdversarialFairness.pdf

The repository is organised as follows:

data/ contains the folders where results and trained models are going to be stored;
src/models contains the implementation of the FAIR models and baselines;
src/utilities.py contains a utilitise for model training and evaluation;
main contains script for starting training and evaluation

Dependencies

The script has been tested running under Python 3..2, with the following packages installed along with the dependencies that are packed in requirements.txt.

In addition, CUDA 11.2 is used.

Conda

Conda environment is available at the environment.yml.

Starting

The code can be started with the

python3 main.py

Reference

If you make advantage of the FAIR model in your research, please cite the following in your manuscript:

@article{petrovic2020fair,
  title={FAIR: Fair Adversarial Instance Re-weighting},
  author={Petrovi{\'c}, Andrija and Nikoli{\'c}, Mladen and Radovanovi{\'c}, Sandro and Deliba{\v{s}}i{\'c}, Boris and Jovanovi{\'c}, Milo{\v{s}}},
  journal={arXiv preprint arXiv:2011.07495},
  year={2020}
}

LICENSE

MIT

andrijaster/FAIR