refining feature importance compute

[thibgo]

Add a specific case where no samples are in the final outcome leaves. In this case, rule importance is set to 0. This avoids a dvision by zero.

[aldro61]

Thanks for the fix @thibgo. However, I don't think that the patch should be done at this level.

For the conjunction case, total_outcome = 0 means that the conjunction never outputs 0. This means that it's composed of rules that do not cover any negative examples. These should simply not be added to the model. In this case, the algorithm should have realized this and just stopped.

What do you think?

[thibgo]

I get the idea and I agree @aldro61 . I will do another pull request with the implementation you suggest.

[aldro61]

Thx! I'm surprised I didn't implement that stopping criterion at the time. I had it in Kover.