Fair Classification via Post-Processing

A post-processing algorithm for fair classification applied to predictors of the form Pr(Y|X) and Pr(A|X), or Pr(A,Y|X), depending on the fairness criterion. Supports (multi-class) statistical parity, equal opportunity, and equalized odds, under attribute aware or attribute blind settings.

See example.ipynb for a quick tutorial. To reproduce our results:

(arXiv 2024 preprint). See the notebooks adult.ipynb, compas.ipynb, acsincome2.ipynb, acsincome5.ipynb, and biasbios.ipynb.
(ICML 2023). Archived under the icml.23 tag; the new version generalizes the old algorithm for attribute-aware statistical parity.

LP solvers. Our algorithm involves solving linear programs, and they are set up in our code using the cvxpy package. For large-scale problems, we recommend the Gurobi optimizer for speed.

Citation

@misc{xian2024UnifiedPostProcessing,
  title         = {{A Unified Post-Processing Framework for Group Fairness}},
  author        = {Xian, Ruicheng and Zhao, Han},
  year          = {2024},
  archiveprefix = {arXiv},
  eprint        = {2405.04025},
  primaryclass  = {cs.LG}
}

@inproceedings{xian2023FairOptimalClassification,
  title     = {{Fair and Optimal Classification via Post-Processing}},
  booktitle = {{Proceedings of the 40th International Conference on Machine Learning}},
  author    = {Xian, Ruicheng and Yin, Lang and Zhao, Han},
  year      = {2023}
}

License

uiuctml/fair-classification

About

Topics

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 1

Uh oh!

Languages