Thanks a lot Scot!
Scot L. Harris wrote:
There is an auto learn mode which does part of the job. From the stuff
I have read and my experience, it is a good idea to run identified spam
as well as ham messages through the sa-learn program to reinforce what
it learns automatically. Also you will need to teach as spam those spam
messages that get through.
Also the bayesian filtering does not actually start until the system has
learned from several hundred messages. And by feeding it a good sized
sample of ham messages as well as identified spam messages you will
reinforce what you believe is spam and ham. It seems to refine the
identification process very well.
I dump identified spam to a holding folder and unflagged spam to
separate folder. Then once a week or so I run sa-learn against those
folders and against my inbox which has all ham messages.
Been doing this for awhile. Currently I get a handful of spam each week
that gets through if that many. Have not had any false positives in a
very long time.
According to the docs feeding it steady diet of ham and spam will keep spamassassin happy and keep up with new tricks that the spammers try.
Well, teaching it the spam that gets thru is difficult in my case, cause there are many users and they download the messages via POP3. Identified spam goes into a seperate folder (and I rotate those folders a few times until they get deleted - like logrotate), so feeding those messages is no problem. Do you have POP3 users on your server ? How do you feed unflagged spam ?
Thanks a lot, Hannes.
