Are params min_cardinality and max_cardinality more self-explanatory?

Are params min_cardinality and max_cardinality more self-explanatory?

Those probably make more sense, and would allow for filtering ranges rather than just below or above a single value.

Since it is closed to what the SuperVectorizer of dirty-cat is doing as heuristic I was wondering if @GaelVaroquaux would have some thoughts regarding the API. Is there a need for adding the minimum cardinality because in practice I would think that low cardinality would be categorical otherwise it will be numerical.

OneHotEncoder can already apply different policies on binary and multi categories features.

So I am just wondering if we need two parameters or a single one?

Any thoughts?

My feeling is that I find the "make_column_selector" API very tedious to use.

I think that there is a benefit to have an object a bit like the SuperVectorizer (maybe calling it differently, that name is a bit ... fun): a simplified API that makes simple things simple.

While SuperVectorizer is great, one reason ColumnTransformer + make_column_selector is tedious is its genericity; the other is that we're trying to configure it with lists of tuples. Having factory-like methods on ColumnTransformer to add new transformers and specify the columns they apply to would make its usage more succinct and pleasant (ColumnTransformer().add(OneHotEncoder(), max_cardinality=5).add(GapEncoder(), min_cardinality=6)), even when not providing something as immediately useful as SuperVectorizer.

So I am just wondering if we need two parameters or a single one?

ColumnTransformer does not route columns of the input exclusively to different transformers, except where remainder is used. Unless we want to add a exclusive param to the ColumnTransformer, the user will need a way to specify both bounds, even to reimplement SuperVectorizer.

@michelkluger Something missing in this PR is related to the documentation. We would need to have this feature demo in the User Guide (next where it is). Then, it would be great to see if there is an example where we could make use of the feature.