Spamassassin rules not firing correctly

[Kai Schaetzl]

Martin Hepworth wrote on Sat, 3 May 2014 12:13:21 +0100:

> You can hand feed corrections via trusyed users, but Ive found that
> autolearn just stuffs up the bayes db

Autolearning is effective in many cases. But you may want to adjust the 
thresholds. e.g. lower especially the ham threshold to something way below 
zero, so misdetected spam "on the brink" isn't autolearned as ham. Also 
very effective is the autolearning of spamtrapped spam. One or two 
accounts that are distributed and surely can only get spam casn already 
have good impact.
I agree that a few users who subscribe just to every list they can find 
and attract shitloads of "legitimate" advertising spam can be a real 
nuisance and may spoil the Bayes DB to an extent that it's not helpful. It 
very much depends on your userbase.
With autolearning one has also to remember that it is not the "raw" hit 
count that gets used. Autolearning discards several rule groups, for 
instance I think it doesn't count the network rules. So, what you think 
should be autolearned because the hits are above threshold may not 
actually hit the required threshold.

Kai

-- 
Get your web at Conactive Internet Services: http://www.conactive.com