predict create erroneous prediction if you use input created from formula (or edit domain or random data)

[jcmhk]

If you use predict with input data created from formula or randomize or with edit domain the result is incorrect.
bug_prediction.zip

I added prompt in source code of predict on my computer:
in function call_predictors of owpredictions.py
it s look like classless_data.domain.attributes are different than predictor.domain in particular on the value of '_compute_value'.
I was unable to adjust this value because I don't know the orange library well enough. I ajusted '_number_of decimal' but it was not the
problem.
I am working on orange 3.35 3.36 and 3.37

[lanzagar]

This is actually not a bug, although it can be very confusing. Maybe there is some additional warning that could guide users to a solution.

What happens is that Formula (and some other transformations) create a new variable that can have the same name, but is not the same object. As you noted, they have a different compute_value.
When Orange tries to use a dataset for predictions it checks that the variables are the same as the data the model was trained on. It can't just match the variables by names, because a normalized columns is not the same as the original one for example.

To tell Orange to forget about transformations (the compute_value) and "reset" a variable (=same effect as saving the data to a file and reading it again), you can use Edit Domain: select the variable and check "Unlink variable from its source variable".
The tooltip provides this explanation:

Make Orange forget that the variable is derived from another.
Use this for instance when you want to consider variables with the same name but from different sources as the same variable.

[jcmhk]

Thank you. Can I submit a pull request to display the warning?

[lanzagar]

Sure. But before you spend too much time preparing the PR, maybe propose a suggestion where and when you would show the warning so we can discuss the solution.