This repository contains the accompanying code for the paper
Paul Gölz, Anson Kahng, and Ariel D. Procaccia: Paradoxes in Fair Machine Learning. 2019.
The paper is freely available at http://paulgoelz.de/papers/equalized.pdf.
This directory contains four scripts:
- implementation.py contains an implementation of the geometric algorithm for finding optimal equalized-odds allocations subject to a cardinality constraint.
- fico.ipynb
is an IPython notebook that allows to replicate our experiments in
Figure 4. It contains instructions on where to obtain the FICO dataset that we
used. If you just want to read the code, the easiest way is to use the link above
to see the preview in your browser, right here on github. If you want to replicate
our results or do your own experiment in our setting, you need to install the
dependencies mentioned below. Then, running
jupyter notebook fico.ipynbopens a browser window, in which you can see our simulation results and easily rerun them. - resmon.py implements the recursion step of the algorithm from Lemma 5, which allows to find equalized-odds allocation with optimal efficiency, while simultaneously satisfying resource monotonicity.
- resmonplot.py generates visualizations for the algorithm in
resmon.pyand was used to construct Figure 3 in the paper. The script is supposed to be called from the command line and contains detailed usage information when called with-h.
We used the following software and libraries in the indicated versions. Newer versions will probably work, but have not been tested.
- Python 3.6.6
- Gurobi 8.0.1 with python bindings
- Matplotlib 2.2.2
- Pandas 0.23.4
- Seaborn 0.9.0
- Jupyter (4.4.0) For academic use, Gurobi provides free licenses at http://www.gurobi.com/academia/for-universities.
To generate the figures in resmonplot.py (and for improved typesetting in the
output of fico.ipynb) an up-to-date version of pdflatex needs to be installed
and be available from matplotlib and the shell.
For questions on the code, please contact Paul Gölz.