[FIX] Naive Bayes: Ignore existing classes in Laplacian smoothing

[janezd]

Issue

As suggested by @lanzagar, issue #2943 can be fixed by ignoring empty classes in Laplacian smoothing.

Plainly, say that y has values male, female, yes and no because y appeared in two different data sets loaded in the same session. Of course, only two values appear in each set. With this PR, only 2 is added to denominators in computation of probabilities, and the probabilities of empty classes are set to 0.

Variable reuse remains a large open issue.

Description of changes

The change also required a minor change in computation of probabilities: instead of computing exp(log(class_probs) + sum(conditional_probs)), the code now computes class_probs * exp(sum(conditional_probls)) because class_probs can now be 0. This should not affect numerical stability.

Includes

• Code changes
• Tests

[codecov]

Codecov Report

Merging #3575 into master will increase coverage by <.01%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3575      +/-   ##
==========================================
+ Coverage   83.97%   83.97%   +<.01%     
==========================================
  Files         370      370              
  Lines       66944    66945       +1     
==========================================
+ Hits        56213    56214       +1     
  Misses      10731    10731

[lanzagar]

Looks good to me.
It still has the [RFC] tag, but unless others have additional comments, I am fine with changing this to [FIX] and merging.