On Sun, Mar 09, 2003 at 10:19:17AM -0500, Joey Hess wrote: > Jonathan Matthews wrote: [snip] > > I'd just like to get some confirmation that these weightings are wrong. > > It's the stock install of SpamAssassin in testing, with no alterations > > made to the config at all. Should I file a bug, change my own > > weightings or go away in shame, having made a fool of myself publicly? > > The deal is that spamassassin's scores are generated using a genetic > algorithm. They "breed" scores against a corpus of known spam and > non-spam, starting with random scores and mutating them up or down, then > seeing how that does and letting the winning mutations thrive. The aim > is to get as few false positives as possible while still catching as > much spam as possible of course. So the scores are not something > hand-tweaked by a human. > > What happens sometimes is it seems that making a score negative reduces > the number of false positives, while not catching any less spam, at > least in their body of spam. And the SA guys, rightly or wrongly, trust > their GA to get it right, and leave these negtive scores in. I have > mixed feelings about this, but it seems to work.
Thanks Joey - I was sure there was /something/ behind it, but I had no idea a GA was used upstream. Seems almost overkill, but if it works ... Think I'll let their best guesses take precedence over mine :-) -- jc -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
