ENH Adds support for drop + handle_unknown=ignore in the OneHotEncoder

[thomasjpfan]

Reference Issues/PRs

Fixes #18072

What does this implement/fix? Explain your changes.

Adds support for the suggestions stated in #18072

1. When an unknown category is encountered and there is a dropped category, then the unknown category will be encoded as the dropped category. (all zeros)
2. The inverse transform will map all zeros to the dropped category if a category was dropped.

[glemaitre]

I think that we should update the docstring of the OneHotEncoder as well.

[glemaitre]

It might be good to show the inverse_transform here.

[glemaitre]

Do you think it would be a good idea to expose an attribute containing the column with unknown categories? I am wondering if the warning will not be too much annoying.

I am thinking that we could have 2 attributes, one containing the column indices and another the unknown categories, when it applies. In this case, we could avoid to warn but you could always check the attributes for sanity check?

[thomasjpfan]

Do you think it would be a good idea to expose an attribute containing the column with unknown categories? I am wondering if the warning will not be too much annoying.

With categories='auto', we would not know if a category is unknown until transform. May you expand on your idea of using the attributes?

If the goal is to avoid warnings, we can hope that the documentation is clear enough and remove the warning.

[glemaitre]

With categories='auto', we would not know if a category is unknown until transform. May you expand on your idea of using the attributes?

Yep this is True. Since it would only be rare, we should not warn so much thought.

[glemaitre]

LGTM

[ogrisel]

LGTM. I would be +0 for having handle_unknown == "ignore" not warn but handle_unknown == "warn" instead, but we can always do that in a latter PR.

In particular @amueller wasn't a big fan of warnings: #18072 (comment)