Modification Bias Score

This is the code for the paper Post-hoc bias scoring is optimal for fair classification

To train the model, run main.py. For testing, run postprocess_dp.py when the fairness constraint is Demographic Parity (DP) or postprocess_eo.py when the fairness constraint is Equalized Odds (EO). These scripts take hyperparameters set in the config.yaml file.

Datasets: Adult Census, COMPAS and CelebA

Dependencies:

• numpy - 1.24.2
• torch - 2.0.0+cu117
• torchvision - 0.15.1
• scikit-learn - 1.3.0
• scipy - 1.11.1
• pandas - 2.0.1
• Pillow - 9.5.0
• PyYAML - 6.0
• tqdm - 4.65.0
• typing_extensions - 4.5.0

Citing the paper (bib)

@inproceedings{chen2024posthoc,
  title = {Post-hoc bias scoring is optimal for fair classification},
  author = {Chen, Wenlong and Klochkov, Yegor and Liu, Yang},
  booktitle = {International Conference on Learning Representations},
  year = {2024}
}