Interesting experiment. I have a couple of questions though: Why use a naive bayes? Is there a personal preference? Have you ever tried to use a SVM with a lin

Interesting experiment. I have a couple of questions though: Why use a naive bayes? Is there a personal preference? Have you ever tried to use a SVM with a linear kernel, which is perfectly suitable for large text corpora? Have you considered removing words with a too low IDF? As far as I see the words with the highest discriminative power are having a positive count of "1", which is not really surprising, that they are highly influencing your results - IMHO, please correct me if I'm wrong. Other than that, interesting read, definitely.