The Impact encoding is not done properly for the validation data for early stopping

[ja-thomas]

We only do this in the training of xgb early stopping, where the factors are already encoded

[Coorsaa]

is this still an issue @ja-thomas ?

[ja-thomas]

yes

[ja-thomas]

I think it might work in the following way when we switch over to CPOs

• We split our dataset before passing it to xgboost.earlystop
• Calculate the preprocs. on the training data and apply the trained CPO on the test data
• Pass the preprocessed test data as a parameter to xgboost.earlystop.

[ja-thomas]

fixed in 385f3cd