Support for `ProbabilisticSet` type in `MLJModelInterface.jl`

[pat-alt]

Is your feature request related to a problem? Please describe.
This is related to this issue and this is more of a proposal than an issue. ConformalPrediction.jl is a new package for uncertainty quantification through Conformal Prediction (CP) for machine learning models trained in MLJ.

Describe the solution you'd like
Conformal classifiers produce set-valued predictions. This is currently not something that is supported for downstream tasks by MLJ. To add conformal classifiers for general use, we would need LearnAPI.jl to support a ProbabilisticSet type (and maybe Set for deterministic classifiers). Of particular interest here is to implement proper evaluation schemes for these types of predictions. A good starting point to this end could be this tutorial from page 11 onwards.

Describe alternatives you've considered
Currently, conformal classifiers in ConformalPrediction.jl are simply subtypes of Supervised (though I might change this to Probabilisitic). They should be more or less ready for general use (except for evaluation, of course, and perhaps other downstream tasks I'm not familiar with).

As I won't get to this myself in the near future, I thought I'd open this up here for discussion. Any thought/comments/contribution very much welcome 🙏🏽

cc @ablaom

[ablaom]

To clarify, LearnAPI.jl is a proposal for a new basement level model interface. MLJModelInterface.jl remains the active API. So @pat-alt's proposal, as I understand it, is to add the following to MLJModelInterface.jl:

abstract type Set <: Supervised end
abstract type ProbabilisticSet <: Supervised end

which sounds good to me.

[pat-alt]

Types have now been added to MLJModelInterface.jl. Evaluation will be addressed in future issues.