Hi,
Jack Gostl wrote:
I've been watching this for awhile, and there is now a pattern to
what I'm seeing.
I'm running a configuration with multiple users sharing a bayes
files. This is an interim move to facilitate the spamassassin
upgrades, and like many interim moves its been going on for a long time.
When I first build the bayes files from my personal folders and my
spam archives, things were great. 99.8% of the spam caught or better.
Then, usually after a week or so, the number starts to drop. Right
now, its down to 97%, in another day or two it will be down below
95%. With the amount of spam we receive, that is a lot of missed junk
mail.
So I blow away my bayes* files, rebuild, and I'm back up to darn near
100% caught. For about a week. Then the deterioration begins again.
Has anyone else encountered this? Is this an artifact of too many
users sharing a spam file?
Also.... I retrain each night, feeding any missed spams plus any new
hams received back through sa-learn. I can't see how that makes it
worse, but who knows.
Do you have autolearn enabled?
Uh... yes? You are suggesting that I turn it off? I had always assumed
that if the Bayes learned something as ham that it shouldn't, sa-learn
was smart enough to undo it.
Change the thresholds for auto learning. Mine are:
bayes_auto_learn_threshold_nonspam -0.1
bayes_auto_learn_threshold_spam 12.0
--
Anthony Peacock
CHIME, Royal Free & University College Medical School
WWW: http://www.chime.ucl.ac.uk/~rmhiajp/
"If you have an apple and I have an apple and we exchange apples
then you and I will still each have one apple. But if you have an
idea and I have an idea and we exchange these ideas, then each of us
will have two ideas." -- George Bernard Shaw