Question on the Over/Under sampling on validation and test splits

[brosscle]

Hi,
I defined a classification hyperpipe that involves a PipelineElement that Oversamples or Undersamples the input dataset. I would like to know if this step is done only on the training split of the nested cross validation or also on the validation and test splits ? Actually, I would like to know if the metrics computed to select the best models and to evaluate them are only computed on the "real" samples and not on the "real + fake" ones (in case of an oversampling), and if it is computed on all the samples and not only the selected ones in case of an undersampling.
Do you know the answer or maybe a document where I can search for the answer ? I have not found it on the documentation but maybe I searched #badly.
Thanks a lot !
Clément

[RLeenings]

Dear Clément,
please excuse the belated answer.
The over and undersampling is SKIPPED in the evaluation for validation and test predictions.
The performance metrics are calculated ONLY on the "real" samples, as you said.
Hope that helps.
Ramona

[brosscle]

Hi, thanks a lot, that's exactly the answer that I was waiting for :)