Reinforcement Learning for Bias Mitigation

This repository hosts the version of the code used for the publication "Algorithmic fairness and bias mitigation for clinical machine learning with deep reinforcement learning".

Dependencies

We have tested this implementation using:

1. Python version 3.6.9 and Tensorflow version 2.6.2 on a linux OS machine.
2. Python version 3.9.2 and Tensorflow version 2.11.0 on a mac OS machine (Big Sur).

To use this branch, you can run the following lines of code:

conda create -n BiasMitigationEnv python==3.7
conda activate BiasMitigationEnv
git clone https://github.com/yangjenny/BiasMitigationRL.git
cd BiasMitigationRL
pip install -e .

Getting Started

To run code:

python BiasMitigationRL/run.py

(UCI Adult dataset automatically loaded for training)

This example uses the UCI Adult dataset, where one is trying to classify income (two classes: <=50K and >50K), and mitigate gender (male vs female) bias. Additional details about the dataset, including all attributes included, can be found here.

After training, performance metrics (auroc,npv,ppv,recall,specificity) and raw prediction results will be saved as csv files in the path. An example run and expected output can be found in example/training_example.ipynb

Citation

If you found our work useful, please consider citing:

Yang, J., Soltan, A. A., Eyre, D. W., & Clifton, D. A. (2023). Algorithmic fairness and bias mitigation for clinical machine learning with deep reinforcement learning. Nature Machine Intelligence, 1-11.