Additional traits for measures

[tlienart]

I'm not a big fan of traits but given we're doing that everywhere we might as well try to be complete? for finite metrics should we maybe add:

• multiclass / binary
• dependent on labelling or not
• robust to imbalance / not

• a docstring like for mljmodels indicating the formula

[ablaom]

Thanks for that.

I think the multiclass/binary distinction is implicit in target_scitype. For example,

julia> info(LossFunctions.ZeroOneLoss())
(name = "ZeroOneLoss",
 target_scitype = AbstractArray{#s56,N} where N where #s56<:Finite{2},
 prediction_type = :probabilistic,
 orientation = :loss,
 reports_each_observation = true,
 is_feature_dependent = false,
 supports_weights = true,)

So, if you parse this, you see target_scitype is AbstractVector{<:Finite{2}}, making this a binary loss function. If the order really mattered (true_positive for example) then you would need to declare AbstractVector{<:OrderedFactor{2}}, so perhaps this also takes care of labelling, for our purposes, no?

The other two traits make sense.

We should also add a measures() method analogous to models() to list the meaures with their traits. For this we probably need a "registry" because Measure subtypes do not include externally defined measures (currently those in LossFunctions).