Can multiple constraints be combined?

[zjiayao]

Hi guys,

Thanks for the fantastics package! It is really intuitive to use and the example notebooks are very helpful to get started.

I have a question -- is it possible to combine multiple constraints (e.g., DemographicParity and EqualizedOdds) together when using say Exponentiated Gradient method? I did some search but was not able to find any documentation on this; is it advisable to subclass fairlearn.reductions.Moment for this purpose?

Thank you!

[MiroDudik]

Thanks for kind words! We currently do not support what you're describing, but technically this is definitely possible.

You could try to create your own moment by combining the code from EqualizedOdds and DemographicParity, but that might not be the easiest route (you'd probably need to extend the functionality of UtilityRateParity to work with events that are not mutually exclusive).

I think technically cleanest would be to subclass fairlearn.reductions.ClassificationMoment into a new class called something like Intersect (or perhaps a better name), which would be constructed by invoking Intersect(moment1, moment2), where moment1 and moment2 would be arbitrarily initialized instances of fairlearn.reductions.ClassificationMoment. You could then write things like:

expgrad = ExponentiatedGradient(
    LogisticRegression(),
    Intersect(EqualizedOdds(difference_bound=0.05), DemographicParity(difference_bound=0.10)
)

[zjiayao]

Thank you, Miro! I'll give that a try :)