On Wed, 26 Feb 2003, Rob Mangiafico wrote: > How is everyone finding the effectiveness of 2.50 vs. 2.43/4 without the > use of a tweaked Bayes filter? Better? Worse? The same?
I've had one false positive (caused by new non-Bayes rules, and not salvaged by Bayes training) and two false negatives, out of 75 spams and ... well, I'm not sure how many hams, something over 100 ... since installing 2.50 and training it. One of the false negatives re-tested as positive after feeding it through sa-learn --spam, so Bayes is definitely making a difference. I trained the Bayes classifier on about 4500 spams, including 83 false negatives from 2.43, and 7000 hams from my mail archives. Of those 83 false negatives, 17 remained false negatives after training 2.50. I then forced the "set 3" scores for BAYES_90 and BAYES_99 to be the same as the "set 2" scores, and found that then only 3 of those 17 remained false negatives. I attribute this to the fact that I don't use Razor/DCC and therefore the "network and Bayes" score set was inaccurate for me. > For our users, they will not be able to effectively send complete emails > into the learning feature of Bayes, so we are not sure how effective > this would be. If they could forward from Outlook emails into it > (without worrying about the headers) that would be great, but I assume > this is of no use? The new feature that causes the original message to be wrapped as a MIME attachment _might_ help you here -- I don't know how thoroughly Outlook (Exchange) dismantles multipart messages. If it actually breaks down the message/rfc822 attachment into the internal Exchange format the same way it does for the top-level message, you're out of luck. ------------------------------------------------------- This SF.net email is sponsored by: Scholarships for Techies! Can't afford IT training? All 2003 ictp students receive scholarships. Get hands-on training in Microsoft, Cisco, Sun, Linux/UNIX, and more. www.ictp.com/training/sourceforge.asp _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
