The original docstring states, that alpha=0 will result in no smoothing. But the Estimators using alpha as a parameter are based on _BaseDiscreteNB, which calls _check_alpha in the fit method. _check_alpha is raising a warning, if np.min(self.alpha) < _ALPHA_MIN and afterwards setting np.maximum(self.alpha, _ALPHA_MIN). Therefore alpha=0 doesn't have any effect and the parameter_constraint can be set to _ALPHA_MIN.

I don't understand this mypy error: https://dev.azure.com/scikit-learn/scikit-learn/_build/results?buildId=43423&view=logs&j=32e2e1bb-a28f-5b18-6cfc-3f01273f5609&t=f370d024-dfa6-536d-2b2d-766ea2a2900c&l=14

When I run mypy sklearn/naive_bayes.py on my local machine, I get Success: no issues found in 1 source file. Can some explain to me, what the issue is?

The original docstring states, that alpha=0 will result in no smoothing. But the Estimators using alpha as a parameter are based on _BaseDiscreteNB, which calls _check_alpha in the fit method. _check_alpha is raising a warning, if np.min(self.alpha) < _ALPHA_MIN and afterwards setting np.maximum(self.alpha, _ALPHA_MIN). Therefore alpha=0 doesn't have any effect and the parameter_constraint can be set to _ALPHA_MIN.

This behavior is also addressed in #22269 and is expected to change soon(ish). I think, for this PR, it would probably be best for you to preserve existing behavior, i.e. allow a user to enter alpha < _ALPHA_MIN and then raise the current warning.

The original docstring states, that alpha=0 will result in no smoothing. But the Estimators using alpha as a parameter are based on _BaseDiscreteNB, which calls _check_alpha in the fit method. _check_alpha is raising a warning, if np.min(self.alpha) < _ALPHA_MIN and afterwards setting np.maximum(self.alpha, _ALPHA_MIN). Therefore alpha=0 doesn't have any effect and the parameter_constraint can be set to _ALPHA_MIN.

This behavior is also addressed in #22269 and is expected to change soon(ish). I think, for this PR, it would probably be best for you to preserve existing behavior, i.e. allow a user to enter alpha < _ALPHA_MIN and then raise the current warning.

Alright, I am going to reverse my changes regarded _ALPHA_MIN. I'll probably get to it tomorrow :)

I revert GaussianNB because an earlier PR already solve the issue: #23583 and it is already approved.

I revert GaussianNB because an earlier PR already solve the issue: #23583 and it is already approved.

Oh, I missed that. I checked the PR section, but I filtered for the ticket number, since the issue told us to put the ticket number in the title. At least we had the same solution :D

I push a small refactoring for alpha and made a proper inheritance for the constraints that are more intuitive.

Ah, that's a good idea. If the constraints of the child class equal the parent class, just use the parent class.

Thanks @Diadochokinetic