[BUG] Error when calling predict_proba with GroupedPredictor using shrinkage and global model about scikit-lego

[FBruzzesi]

Hey @cerlymarco , as you can see this issue was mentioned in a few places. Just to add some description/context to each one of those:

• GroupedPredictor refactoring was a first attempt to fix the issue reported here, as well as expanding the functionalities. Since this was getting a bit messy we decided to follow another approach, which led us to the next two points.
• GroupedPredictor patch follows your suggestion to raise an error if shrinkage is used in combination with a classification task. This is a somewhat breaking change which is already in the main branch.
• In HierarchicalPredictor and HierarchicalTransformer we set up the discussion for how to deal with that as well as adding other features to meta estimators that are somewhat similar to Grouped* ones.
  • I don't want to make promises, but you could expect HierarchicalPredictor (or even better a HierarchicalClassifier) that works with classification tasks and shrinkage to come fairly soon
  • Available since 0.8.0 release

[cerlymarco]

Hi @FBruzzesi, thanks for your reply! That seems reasonable. If that is the case, an error/warning may be raised when a classifier (an estimator with predict_proba method) is provided with shrinkage enabled. Otherwise, a proper shrinkage method for probabilities could be implemented

[FBruzzesi]

Hey @cerlymarco, thanks for reporting this bug. I can replicate the issue and will take a closer look.

My first impression is that shrinkage was intended for regression problems only, as running .predict(...) with the same data you provided will yield values that are not classes and that shouldn't be the expected output.