This message was deleted.

Hey. You can compute the features with preprocess, manually train the model and then use it to compute predictions like here or you could also leverage the fact that lightgbm automatically uses columns with categorical data type as categorical features using a pipeline like here.

Perfect, thank you!