On 27 Feb 2002, Joey Hess <[EMAIL PROTECTED]> wrote: > Daniel Rogers wrote: > > LINE_OF_YELLING seems to have jumped from a score of 0.70 in SA 2.01 to a > > score of 5.442 in SA 2.1. This strikes me as rather a lot.
(I've only been using SA for a little while, so you can take this with a bit of salt.) It really does seem to me like the weightings are less useful in 2.1 than in 2.01. Some of them are clearly too large, and some of the negative weights indicate that the algorithm is "overtrained" and looking for patterns when there is not sufficient evidence to decide. It seems like human intelligence in assigning weights would still be useful. Perhaps rather than the GA running all by itself to decide weights, it would be better to have them set by humans, and then applied to the spam and nonspam corpus to see if the settings are useful? SpamAssassin looks really cool, but at the moment I've had to go back to 2.01 to feel comfortable that I won't get incorrect results. -- Martin _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
