improvements in naive_bayes

[fabianp]

• across the whole module estimated parameters do not end with underscore
  • non-standard names are used for some variables, like unique_y (called classes in SGD and labels in SVM), we should think about unifying these.

[larsmans]

I'd go for labels. By the estimated parameters, do you mean proba_y, theta and sigma? Wouldn't changing those break pickle'd GNB instances?

[larsmans]

Commit d7e64be introduced some changes to Gaussian Naive Bayes, but more work remains to be done. It would be elegant to reformulate GNB as a linear classifier, just like Multinomial NB.

[amueller]

@fabianp do you think the issue is resolved? Or is there something else to do?

[amueller]

The unique_y was also fixed at some point, classes_ is now used everywhere as far as I can tell. I hope #890 will help unify things even more.