Allow dropping of first level in OneHotEncoder

[CameronBieganek]

When dropping a level of a categorical variable for one-hot encoding, R and Python both default to dropping the first level. It would be nice to have that option with OneHotEncoder in MLJ. For example, I have a data set with a bunch of columns that are coded like this:

Code	Meaning
1	Yes
2	No
9	Missing

In this case I would prefer to drop the first level rather than the "missing" level. (Actually, I would really prefer to drop the "No" level...)

[ablaom]

How about we add a drop_first option and a drop_level option. So, eg, drop_level=9 or drop_level="No"?

[CameronBieganek]

That sounds good! Although the drop_level case is a little tricky, since you might want to be able to control the dropped level on a per column basis. Scikit-learn allows the drop argument to be a vector, though I think a dictionary or array of pairs would probably work better than a vector. (There could be columns with continuous variables in between the categorical columns...)

[ablaom]

since you might want to be able to control the dropped level on a per column basis.

Right! 😳

I'd vote for the dictionary.