`HierarchicalPredictor` and `HierarchicalTransformer` about scikit-lego

[koaning]

Shrinkage doesn't really make sense, or at least not for all transformers. In fact the current GroupedTransformer doesn't support it.

Agree!

Should the group columns be returned as they are from the .transform(X) operation?

It feels like it would be fine to default to not returning the group columns. Mainly because users could select them seperately in the pipeline if they were so inclined. Might make for a good parameter though.

Should the .transform(X) output maintain the input type? Hence if pandas dataframe maintain the index as well?

I would prefer not to rely on anything pandas specific with the advent of possible polars support. Is there a use-case you had in mind where that's required?

[FBruzzesi]

Ok, these will definitely be (at least) two separate PRs, one for predictor and one for transformer (the predictor one could land late this week 😁)

Might make for a good parameter though

"passthrough" vs "drop" seems suitable

Is there a use-case you had in mind where that's required?

Absolutely not, I started to rely on indexes as less as possible, but since pandas and polars input/output are supported by scikit-learn, I am wondering what is the behavior one should expect.