Basically do some sort of penalty or regularizer to avoid overfitting. We only do train data-target encoding and then follow or map the same thing in testing as per the train (there is no target in test data ).