I don't think the section on automatic foldering is convincing. It, once again, comes to the conclusion that it doesn't work/Naive Bayes isn't accurate enough, yet POPFile shows very high accuracy for a wide variety of users (this is from real usage data).<p>It's odd that his thesis mentions POPFile and then doesn't go on to actually test it. He also uses the Enron data set for many of his tests which has very little header data in it.