Who else is using the Bayesian classifier from the current 2.50-cvs? What kind of results is it giving you?
I fed the learner about 2700 spams and 5600 non-spams, taking special care to give it all the false-negatives I've been squirreling away. I then re-checked 169 of those same false negatives (happened to be the contents of one spam mailbox); 31 messages scored over 9 because of non-Bayes rule changes since 2.40, so I ignored those. Of the remaining 138: 63 were still NOT classified as spam (45.7%). 75 were correctly classified as spam (54.3%). Not bad, but not as good as I'd hoped. No message got less than BAYES_60, which I expected since they were messages it learned on. 7 messages scored ONLY on BAYES_* rules, but of course even BAYES_99 (two messages scored only on that) isn't worth 5 points. Since then SA has processed 72 new spam messages and a couple of dozen non-spam (I don't keep count of those), but none where the BAYES_* rule caused the total score to cross the 5-point boundary in either direction. It has sent a few slightly-above-zero scores into the negatives for the non-spam, but that doesn't really help. Just curious how this compares to what others are seeing. ------------------------------------------------------- This sf.net email is sponsored by:ThinkGeek Welcome to geek heaven. http://thinkgeek.com/sf _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
