Unsupervised Adaptation for Fairness under Covariate Shift

This repository serves to open-source the code used in the paper: "Fairness under Covariate Shift: Improving Fairness-Accuracy tradeoff with few Unlabeled Test Samples ".

Preliminaries

Getting started

To avoid any conflict with your existing Python setup, it is suggested to work in a virtual environment with virtualenv. To install virtualenv:

pip install --upgrade virtualenv

Create a virtual environment, activate it and install the requirements in requirements.txt.

virtualenv env
source env/bin/activate
pip install -r requirements.txt

Running the code

Generating Data Splits

python generate_data_splits.py --ds_name {0} --gamma {1}

Here
{0}: is the dataset name - "arrhythmia", "adult", "communities", "drug"
{2}: gamma - the shift factor for the test set

Similarly, the following command can be used to generate the asymmetric splits

python generate_asymm_splits.py --ds_name {0}

"gamma" values for the groups can be adjusted in "create_asymm_covar_shift" function

Running the main file

To run the model

python main.py --model_type {0} --ds_name {1} --do_train --k_shot {2}

Here
{0}: is the model name - can be assigned to anything of interest, eg - "wass_and_entropy_model"
{1}: dataset name
{2}: the number of unlabelled test samples utilized
The remaining arguments can be found in "arguments.py" file with explanations

Disclaimer

This is not an officially supported Google product.

Citation

If you find our work relevant, please cite as:

@article{havaldar2023improving,
  title={Improving Fairness-Accuracy tradeoff with few Test Samples under Covariate Shift},
  author={Havaldar, Shreyas and Chauhan, Jatin and Shanmugam, Karthikeyan and Nandy, Jay and Raghuveer, Aravindan},
  journal={arXiv preprint arXiv:2310.07535},
  year={2023}
}