I don't think logistic regression would necessarily be better than naive bayes for this. It is not you can only use naive bayes for only a bag-of-words model. E

I don't think logistic regression would necessarily be better than naive bayes for this. It is not you can only use naive bayes for only a bag-of-words model. Either are ML algorithms that you can use with any features you want. However I am sure they tested both. Also, I think they should have used a different algorithm for the global model, which gives a better performance at the expense of more training training.