Has anyone been able to reproduce the experiments?

[cmougan]

I am trying to reproduce the disparate mistreatment experiments. There are no guidelines, I am missing:

• Python version. To reproduce the envieronment.
• Python packages. To install with pip
• Docstrings.
• Some minimal code structure that looks fairly similar to pandas or scikit-learn.

After several attempts and random guessing the above issues, the code still does not runs. It throws errors.

Any help will be very much appreciated.

Many thanks :)

[cmougan]

@mbilalzafar is it possible to know the python version and the requirements.txt used?

[mbilalzafar]

Hi @cmougan. Thanks for you interest in our work :)

We used python 2.7 (we developed the code some while ago). The following environment setup should work:

conda create -n py27 python=2.7
conda activate py27

Then install the dependencies using the following requirements file:

numpy==1.16.6
scipy==1.2.3
matplotlib==2.0.2
cvxcanon==0.1.1
ecos==2.0.4
scs==1.2.6
cvxpy==0.4.9
dccp==0.1.6
sklearn
pandas

The demo files mentioned in disparate_mistreatment/README.md (synthetic_data_demo/decision_boundary_demo.py, synthetic_data_demo/fairness_acc_tradeoff.py and propublica_compas_data_demo/demo_constraints.py) should provide minimal code examples for training unconstrained and fairness-constrained classifiers.

The main function being used is train_model_disp_mist in fair_classification/funcs_disp_mist.py and it contains documentation of the parameters.

Let me know if you have any further questions.