Support has_time parameter in CatBoostEncoder

[kylegilde]

Expected Behavior

In CatBoost, the highly-cardinal categorical features can be encoded in two ways using the has_time parameter:

1. w/ random permutation (has_time=False)
2. w/o random permutation (has_time=True)

https://catboost.ai/en/docs/concepts/parameter-tuning#internal-dataset-order https://github.com/catboost/catboost/issues/1076#issuecomment-905039708

Actual Behavior

The CatBoostEncoder only supports the 1st method.

Proposal

Add the has_time parameter to CatBoostEncoder in order to support ordered datasets

[glevv]

I think it's the opposite: existing implementation of CBE uses no permutation, so it is time-aware implementation. Old implementation used LOO once, so it was 1 round of permutations.

Maybe original CBE implementation coders could correct me if I'm wrong.

[kylegilde]

@GLevV , I wasn't notified when you made your comment.

Interestingly, the docstring indicates that the random permutation is occurring, but I don't see that it is happening anywhere in the code itself.

@PaulWestenthanner, I wonder if the docstring is incorrect?

[PaulWestenthanner]

I'm not too familiar with out implementation of catboost encoder. I'd need to have a look at it. I cannot guarantee the docstrings are correct