Am 26.09.2014 um 18:19 schrieb John Hardin: > On Fri, 26 Sep 2014, Reindl Harald wrote: > >> frankly the biggest problem is the large amount of idiots hit >> a "spam" button whenever they can to stop receive some sort of >> mail - i had that even in my own family "can't you block that?" >> followed by "yes" after asking "have you subscribed there?" >> >> guess how likely it happens for a large mail provider sending >> 100% clean mail with double-optin to 10000 persons when there >> are 5% complete idiots - it results in 500 spam reports about >> the same message - the hardest job by maintaining a blacklist >> is to catch that idiots and prevent harm for innocent > > This is exactly why user-submitted training messages should be manually > reviewed before being learned
i know - the 3 messages i talked also where in fact no spam but composed that stupid to match a lot of common spam classifiers and so neutralized them in bayes which was enough to get spam scored 0.2 points under the milter-reject value the current bayes contains 3466 messages all reviewed by me it showed me even review is not error safe hence the word "autolearn" is a no-go for me - i doubt that a software learning by it's own decisions will do it always right i saw that multiple times with a "Barracuda Spamfirewall" also using SA behind the scenes with autolearning - after a new deployment all was just fine, some manual training and perfect results over the moths it became every time worser * more junk slipped through * more and more legit mail got tagged
signature.asc
Description: OpenPGP digital signature
