Another question with respect to training decision trees - do we really need to encode categoricals for random forest of for that matter any decision tree. Or do we get some computational benefit from using ordinal encoders.
rgds
RandomForest expects only numerical values. If one would try to input the X matrix without encoding, then it will raise an error.
In the future, scikit-learn might get to a point where categorical features could be natively supported with some inner encoding when necessary, but this is not implemented yet.