Should PipeOpLearner pass on a `Prediction` during train

[pfistfl]

I know we talked about this, but it is really odd/annoying to have to explain to a user that we require learner_cv instead of learner every time I want to do any post-processing of predictions.

Example:

• Fairness -> I want to learn thresholds / flipping probabilities during train.
• Threshold tuning
• Any prediction post-processing that is learnt.

[mb706]

Current types:

Learner: (Task|Task) -> (NULL|Prediction) LearnerCV: (Task|Task) -> (Task|Task)

ThresholdTuning (using only prediction): (NULL|Prediction) -> (NULL|Prediction) ThresholdTuning (using CV data, proposition 1): (Prediction | Prediction) -> (**** | Prediction) ThresholdTuning (using CV data, proposition 2) ist vll ein Learner: (Task|Task) -> (NULL|Prediction)

IMHO "proposition 2" makes sense, since threshold tuning trains a decision stump on one feature, so it makes sense to have predictions as a feature.

[mb706]

Decision: We currently agree to keep the current problem; maybe we have to think about communicating this better or at least create good gallery posts for people to copy-paste from.