Popfile Update #4

I cut down the number of "buckets" into which messages are classified, and Popfile’s accuracy in filtering spam has gone up to about 95%. However, I still get false positives, roughly one per day. Classifying real mail as spam is much worse than letting a few spams through. This is an area where Bayesian filters like Popfile do better than rule-based systems, but better may not be good enough. After spending hours training Popfile on over 3,000 messages, I’m at the point where I think it’s worth using, though just barely.