Preprocessing: difference in preprocessing pipelines

[ajdapretnar]

• [ ] What's wrong?

  We usually teach both of these approaches are valid.

  

But actually the first workflow overrides the default preprocessing of the learner, which causes errors sometimes.

• [ ] How can we reproduce the problem?

Try heart-disease with Logistic Regression. It fails and the error is unclear for a beginner (ValueError: Wrapped scikit-learn methods do not support multinomial variables.). Adding Continuize solves the problem, but this is not clear from the error.

Proposed solutions:

• Have a nicer error that tells the user what to do / what the problem is.

• Write this in the docs.

• Handle the difference (make both workflows behave the same). Should we?

• [ ] What's your environment?

• Operating system: Big Sur

• Orange version: 3.29.dev

• How you installed Orange: git

[janezd]

Write this in the docs.

... in Preprocess and in Test and Score.