noise_fairlearn

Introduction

This repo consists of code used for reproducing the results shown in the paper Noise-tolerant fair classification (Neurips 2019) by Alexandre Louis Lamy, Ziyuan Zhong, Aditya Krishna Menon, Nakul Verma. It implements a general interface for noisy fair binary classification. It currently supports the following:

fairness criteria:

• Demographic Parity
• Equality of Opportunity

classifiers:

• Algorithm from A Reductions Approach to Fair Classification

datasets:

• UCI Adult
• UCI German
• UCI Bank
• COMPAS
• Law School

Usage

Directly use demo.ipynb or run

python3 run_experiment.py --eval_objective test_tau --dataset compas --rho-plus 0.2 --rho-minus 0.2 --frac 1 --criteria DP --classifier Agarwal --trials 3 --plot-result

The details of parameters can be found in the definition of the experiment function in util.py.

Reference

Code is modified on top of code in the following repos:

files in fairlearn are modified from:

https://github.com/Microsoft/fairlearn

(not used in the paper) files in fair_classification are modified from:

https://github.com/mbilalzafar/fair-classification

(not used in the paper) files in fairERM are modified from: https://github.com/jmikko/fair_ERM