`step_lencode()` but like `step_dummy_multi_()`

[EmilHvitfeldt]

[jrosell]

Should we have a parameter to be able to tune and control the smoothing strength in the target encoding?

[EmilHvitfeldt]

definitely

[jrosell]

I saw that you have already implemented smooth = TRUE parameter in https://embed.tidymodels.org/dev/reference/step_lencode.html for https://github.com/tidymodels/embed/releases/tag/v1.2.0, but not yet the strength. I understand that we now have the equivalent of smooth="auto" in scikit-learn.

I would really like to have a new numeric parameter to adjust the strength beyond the empirical Bayes estimate. Is something that would be in the {embed} package or it's complex because of cv?

[EmilHvitfeldt]

I'm not opposed to having a strength argument. However for this first method (needed for conf) i wanted to go with what i felt is the better solution.

IMO it is hard to find a good default for strength so having a method that requires tuning seems less ideal. But will be added for completeness at one point

[jrosell]

Oh. I understand. Thanks for your reply.