Easily retrieve mapping from OrdinalEncoder

[oasidorshin]

Describe the workflow you want to enable

It would be nice to be able to easily retrieve mapping in the form of a dictionary

"category_a": 0,
"category_b": 1,
"category_infrequent": 2,
...

Currently .categories_ attribute only retrieves list of seen categories, without mapping.

This becomes especially important with options to handle missing or infrequent values, which are leading to questions "What value does infrequent categories map to?" and so on.

Describe your proposed solution

Add .categories_map_ attribute to OrdinalEncoder

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[ogrisel]

It's possible to construct the category maps via list and dict comprehensions:

category_maps = [
    {cat_name: cat_idx for cat_idx, cat_name in enumerate(cat_list)}
    for cat_list in oe.categories_
]

or as a dict of dicts to leverage feature names when available:

category_maps = {
    feature_name: {cat_name: cat_idx for cat_idx, cat_name in enumerate(cat_list)}
    for feature_name, cat_list in zip(oe.feature_names_in_, oe.categories_)
}

Maybe we could expose a public property (or two, in the presence of the feature names) on the OrdinalEncoder but I also worry about the magic aspect of properties.

Precomputing those reverse mappings is also possible but redundant.

Another alternative would to use methods (instead of properties) to generate those reverse mappings on depend, maybe with optional argument to do it only for a specific set of features.

[ogrisel]

This becomes especially important with options to handle missing or infrequent values, which are leading to questions "What value does infrequent categories map to?" and so on.

I am not sure how providing encoding maps would help answer that question: what would be the "key" in the dict for infrequent or unseen categories?

But we could expose dedicated fitted attributes to make the resulting encoding of missing and infrequent/unseen categories explicit for each feature.

EDIT: such an attribute already exists for infrequent categories: OrdinalEncoder.infrequent_categories_

[oasidorshin]

@ogrisel

Thank you for your comments!

I will try to argue more about my motivation for each question:

1. Possible to construct the mapping by hand - my main gripe with this approach is that the assumption "order of categories in this list == encoded labels" is currently not outlined in the docs. So the solution would be to either outline encoding algorithm (inc. missing/unseen values) it in the docs, or to give method to produce exact mapping (which would still be very useful in reducing boilerplate code in downstream applications)

2.1. What would be the "key" for unseen categories - I think it is either some predefined "_unseen_category_hope_not_in_user_data", which is ugly, or possibly some new attribute like .unseen_encoded_value

2.2. OrdinalEncoder.infrequent_categories_ - while it is very useful for checking what categories turned out to be infrequent, it still doesnt answer the question into which value they are being encoded.