How is this going, @nilichen?

I managed to get fit_transform and inv_transform working for general cases, but got stuck in deciding how handle_unknown should interact with different options, especially drop and handle_unkown (e.g., whether NaN should be treated as another category or not depends on other options). And the logic in the code starts to get obscure.

My personal view is that this function has become a bit too complicated and it might be more straightforward that users deal with NaN/None in pandas as usual. Otherwise, #13028 might be a better option to move forward.

how handle_unknown should interact with different options, especially drop and handle_unkown (e.g., whether NaN should be treated as another category or not depends on other options). And the logic in the code starts to get obscure.

Thanks for sharing your experience with it @nilichen . Long term maintainability of these options is certainly a concern, particularly if we want to add other encoders (that support them) in the future.