Sven Hartge wrote:
> Um 15:20 Uhr am 26.06.05 schrieb Marc Haber:
>>Or is bayes tagging so advanced in these days that you can easily use
>>autolearning for the entire system's mail traffic and still have decent
>>results?
[ Snipped statement about bad per user results ]
> After switching to only one global bayes database (and after consulting
> the data securty official ) the hit rate was greatly improved.
While this is, in the strictest sense, a Very Bad Thing(tm) since each
user's mail is going to have a different set of words and phrases to learn I,
too, set SA's bayes to global and leave it there. I think that unless one
gets a huge amount of mail 1 user per DB might not be the best and a few
(maybe up to a dozen) users is acceptable.
Besides, with a decent enough margin of error on the system filter there's
nothing preventing the user from running their own bayes on just their mail.
Personally my system uses exim/sa-exim/sa and rejects at quite a low number
(7.5?). Even so I have Thunderbird's bayes turned on and it catches stuff
that SA lets slip by. This is mainly because SA's bayes is consdiered "just
another test" and so bayes of 100% but AWL, ALL_TRUSTED and a few other
positive tests can get obvious spam (to Thunderbird) past SA as a marginal
message.
--
Steve C. Lamb | I'm your priest, I'm your shrink, I'm your
PGP Key: 8B6E99C5 | main connection to the switchboard of souls.
-------------------------------+---------------------------------------------
