Skip to content

sujeongcha/bandit-under-fairness-constraints

 
 

Repository files navigation

Sentencing Under Fairness Constraints: Finding a Fair Policy with Offline Contextual Bandit

Final Report

For details about our experiments on the NODA dataset, please refer to here.

Disclaimer

Reproduction can only be done using ProPublica (located in dataset/propublica) as NODA cannot be published for confidential reasons. Our code for data processing on NODA cannot be shared for the same reason.

Installation

This code has been tested on Ubuntu.

First, install Python 3.x, Numpy (1.16+), and Cython (0.29+).

The remaining dependencies can be installed by executing the following command from the Python directory of :

pip install -r requirements.txt

Usage

The experiments presented in the report can be executed by running the following line in the command:

 python -m experiments.bandit.recidivism recidivism_all --n_trials 5 --definition GroupFairness --e 0.1 \
 --d 0.05 --ci_type ttest --n_iters 2000 --n_jobs 15  --r_train_v_test 0.4 --r_cand_v_safe 0.4 --rwd_recid -1.0 \
 --rwd_nonrecid 1.0 --use_score_text --data_pct 1.0 --add_info all
  • recidivism_all: folder name to save the results
  • --add_info: covariate sets to use ('none', 'all', 'judge', 'screen_ada', 'trial_ada') - only applicable to NODA dataset

Reading Experiment Outputs

Model output comes in .h5 file format. reading robinhood output.ipynb in ipynb folder will transform the result to a well-formated dataframe. We provide recidivism_judge.h5 as an sample output file.

Acknowledgement

Original code comes from https://github.com/sgiguere/RobinHood-NeurIPS-2019, and it has been modified to be applicable for our experiments on NODA.

License

Code for RobinHood is released under the MIT license, with the exception of the code for FairMachineLearning (located in Python/baselines/fairml.py), which are released under their licences assigned by their respective authors.

About

Implementation of safe offline bandit algorithms.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 58.0%
  • Jupyter Notebook 42.0%