This project contains an accessible implementation of the OTF cost function proposed in the paper Optimal Transport of Classifiers to Fairness published at NeurIPS 2022.
The OTF cost projects a probability distribution to the closest distribution in the set of all fair distributions, where closeness is defined in terms of Optimal Transport cost. As such, OTF quantifies the unfairness of a model while taking the input features of individuals into account.
An example use of the OTF method for the Adult dataset is given in main.py . The actual implementation of the cost is in otf.otf_cost, where a fairness notion as given in otf.linear_fairness_notion is expected. In otf.predictor, a generic probabilistic model is implemented that uses the OTF cost as an additional cost term to optimize during training. Finally, otf.evaluation computes some metrics as explained in the paper.
When using the OTF cost, please make sure to tune the reg_strength hyperparameter at the very least, e.g. in the range [0.01, 0.001, 0.0001]. Also, the computation of the OTF cost can be sped up by reducing the nb_epochs parameter and increasing the margin_tol and constraint_tol.
If you found our code useful, please cite our paper:
@inproceedings{buyl2022otf,
title = {Optimal Transport of Classifiers to Fairness},
author = {Buyl, Maarten and De Bie, Tijl},
booktitle = {Advances in Neural Information Processing Systems},
volume = {35},
pages = {33728--33740},
year = {2022}
}