OneHotEncoding should retain metadata

[rofinn]

The current OneHotEncoding transform seems to always just return a NxP binary matrix despite knowing what the categories are and potentially being passed a type that would allow you to retain that info.

For example, if I have OneHotEncoding(Hour(0):Hour(1):Hour(23)) and then I pass a dataframe with an HoD column, I could easily see that transform returning a new dataframe with a column name for each hour. Similarly, if we took a function argument, we could do something like:

FeatureTransforms.ohe(Hour, data; dims=:time)

This would return a new KeyedArray or AxisArray that retains the category information for the p dimension, and also uses the dims argument to mean use dimension keys for KeyedArrays or AxisArrays.

[glennmoy]

Can you explain what you had in mind for the function argument? I'm not sure I see how it translates to keeping the category information? Is it that the function determines what the category labels are?

[rofinn]

Is it that the function determines what the category labels are?

Yes, so if I have an axis with datetimes I might use hour to lazily generate the categories. For AxisArrays and KeyedArrays we could then return n (:time/DateTime) x p (:category/Hour).