TensorFlow implementation for the paper Learning Controllable Fair Representations, AISTATS 2019.

Requirements:

• Tensorflow
• pandas
• numpy
• tf_utils

git clone [email protected]:jiamings/tf_utils.git
cd tf_utils
pip install -e .

python -m exmaples.adult

python -m examples.adult --lag

If MIFR then the e hyperparameter values corresponds to individual lambda parameters, if L-MIFR then they correspond to epsilon contraints in the paper.

• e1: Upper bound for MI
• e2: Adversarial approximation to Demographic parity
• e4: Adversarial approximation to Equalized odds
• e5: Adversarial approximation to Equalized opportunity
• disc: discriminator iterations

If you find the idea or code useful for your research, please consider citing our paper:

@article{song2019learning,
  title={Learning Controllable Fair Representations},
  author={Song, Jiaming and and Grover, Aditya and Zhao, Shengjia and Ermon, Stefano},
  journal={arXiv preprint arXiv:1812.04218},
  year={2018}
}

utils/logger.py is based on an implementation in OpenAI Baselines.

tsong [at] cs [dot] stanford [dot] edu