This repository provides the implementation for an algorithm that aims to reduce bias in classification.

The associated paper is the following :

Classification with Fairness Constraints: A Meta-Algorithm with Provable Guarantees : https://arxiv.org/abs/1806.06055

To run the algorithm, clone the repository and follow the given steps.

In this repository, we have implemented 3 algorithms. The difference between the algorithms is the fairness function they want to optimize.

The fairness functions considered here are :

• Statistical parity
• False Discovery
• Statistical parity + False Discovery

To get a better idea of the algorithm and the constraint functions, we refer the reader to the paper.

The algorithms can be run on the pre-processed datasets provided (Adult and Compass). To run any of the above algorithms, simply run the corresponding file with the data folder as the argument :

    $ python3 FalseDiscovery.py ../Data/data

This will run the algorithm for input fairness parameter 0.1, 0.2, ..., 1, and for each case output the accuracy over the test data and the observed fairness values.

Fairness Comparison

We also provide another way to run the algorithm and compare it with other fair-classification algorithms. To do this, use the framework provided by https://github.com/algofairness/fairness-comparison/blob/master/README.md. The details of the implementation are provided in the FairnessComparison folder.

Adding a new algorithm

To add a new algorithm based on a different constraint, simply extend the General.py file, and implement the functions for optimization based on the constraint function chosen.