Right now, we cannot deal with this when using categorical default rules, as they will render the categorical features unsupported rather than ignored.
While I am not aware of models that benefit from both unencoded categorical features and normalisation, we could still consider adding this, especially since users users are likely to apply unnecessary normalisation to, for example, tree-based models.
Proposal:
- make
add_categorical_default_rules accept not only bool but also an enum CategoricalRule with items UNSUPPORTED, IGNORED, NONE.
- handle it appropriately
- pass
add_categorical_rules through to further subclasses, in particular TakeColumns.
@schroedk FYI