The biggest issue I see in this generic approach is, well, that it’s generic. 😅 If we go down this path, it emphasizes that fairness can be seen as an intrinsic problem of the model, rather than being a sociotechnical one.
On the other hand, it’s a lot better than using an inappropriate data set. So it might be a good alternative until we find a better one.
What do you think, does this make any sense?
I’d be okay with gender queer as a term, as it makes clear the intention is to represent people outside a binary gender scheme. I can’t speak for all non-binary though. 😬
@Roman that’s actually a good idea. Without having read the original papers, I wouldn’t have understood it either. Having a approachable theoretical section is a unique selling point for the fairlearn documentation I guess.
Re. collecting gender demographics, I have recently chatted with Queer in AI folks, and they suggest the following:
Gender: (please select the options which are applicable to you)
 Non-binary / Genderqueer / Third gender
 Genderfluid / Gender non-conforming
 Prefer not to say
 Specify your own (open text box)
The key aspect is that this is a multi-select. When this kind of data is collected, there should be an explanation that the data is collected to audit for fairness and/or mitigate any fairness issues, and that if you pick "prefer not to say" you will not be included in any fairness audits.
I also like these two sources:
Respectful Collection of Demographic Data | by Sarai Rosenberg | Managing on the Margins | Medium
In scikit repo
conf.py which you pointed in description of the issue:
I am confused what should we put for in the case of
dependencies - requirements.txt. If you look at scikit-repo requirements - https://github.com/scikit-learn/scikit-learn/blob/309f135c3284d7db6e23ca81a87948c7066a3949/doc/binder/requirements.txt, it's confusing.
What should we put in the requirements.txt in case of fairlearn repository?
predict_proba. Note that our meta estimators like
ExponentiatedGradientdo NOT produce probabilities the same way, so we've very consciously chosen not to name it
Assessment: our user guide covers both classification and regression metrics https://fairlearn.org/main/user_guide/assessment.html#metrics
Mitigation: The table shows which techniques should work for regression: https://fairlearn.org/main/user_guide/mitigation.html (everything but ThresholdOptimizer)
Admittedly, the user guide isn't super comprehensive yet and has some gaps, but there is a regression section for fairness constraints that can be plugged into our
fairlearn.reductions techniques: https://fairlearn.org/main/user_guide/mitigation.html#fairness-constraints-for-regression
We don't have examples with regression mitigation yet, but this test case code might be useful to you: https://github.com/fairlearn/fairlearn/blob/62fc80c77bcd3bef6a3d7bc44e54827ec9fb8d09/test/unit/reductions/grid_search/test_grid_search_regression.py#L64
If you're willing to explain your use case we could try to advise on how to use Fairlearn. Let us know!