Counterfactual Fairness through Data Preprocessing

We implement the Fair Learning through dAta Preprocessing (FLAP) algorithm to learn counterfactual fair decision through data preprocessing and the Conditional Distance Correlation test (Wang et al., 2015) to detect discrimination.

Requirement

python >= 3.7

Usage

Fair learning

Admission example (Wang et al., 2019) with sample size 5000: First run main.py with the data config file to generate the evaluation metrics

python ./main.py -c ./config/admission_n5000_eval_AllDiffScoreChange.json

Then run plot.py with the same config file to generate charts with previously generated metrics

python ./plot.py -c ./config/admission_n5000_eval_AllDiffScoreChange.json

Fairness test

Admission example (Wang et al., 2019): First run main.py with config files for different sample sizes

python ./main.py -c ./config/admission_n50_test_AllChange.json
python ./main.py -c ./config/admission_n100_test_AllChange.json
python ./main.py -c ./config/admission_n200_test_AllChange.json

Then run plot.py with the config file for the biggest sample size to plot the results for all sample sizes

python ./plot.py -c ./config/admission_n200_test_AllChange.json

Config files for other examples can be found in the config/ folder. The configs for fairness learning evaluation have _eval_ in their names, the configs for fairness test have _test_ in their names.

Real data analyses

We provide ipython notebooks for real data analyses

• Adult income data: the data is available at the UCI Machine Learning Repository, the analysis in real-data-adult.ipynb can be reproduced after downloading the adult.data and adult.test files and storing them under data/Adult/ folder.
• COMPAS recidivism data: the data is available at the ProPublica Repository, the analysis in real-data-compas.ipynb can be reproduced after downloading the compas-scores-two-years.csv file and storing it under data/COMPAS/ folder.
• Fintech data: the analysis is in real-data-fintech.ipynb.