I’ve found that many installations of SpamAssassin work well, but end-users don’t optimize them to train their Bayesian filters on false negatives (i.e., spams that get through). A LearnAsSpam IMAP folder is a great solution for this, which even works with Exchange users (they just need IMAP turned on at the server).
I wrote up directions for installing this on the SpamAssassin wiki.
Manuel | 10-Feb-10 at 4:49 pm | Permalink
I used a similar solution for the past years, but now I use to train my bayesian filters with the help of custom IMAP flags.
Since Mozilla Thunderbird uses the flags Junk and NonJunk for marking mail messages as spam or ham, a newly created perl script traverses over all user’s mailboxes in order to watch out for such flagged messages. Those are then leaned either as spam or ham and even flagged as LearnedJunk and LearnedNonJunk.
Additionally you can create new tags in Thunderbird itself (called LearnedJunk and LearnedNonJunk) to colorize previously learned messages.
As soon a new mail arrives and Thunderbird marks it as junk automatically or you do so by hand, the mail will be learned as spam and colorized. If Thunderbird thinks a mail is spam and you disagree and unmark it, the mail will be re-learned as ham.
This posting is really old, but might be still of interest for some users, since there seems to be a huge lack of information concerning training spamassassin in a comfortable way. Some might also check out my german blog posting and the corresponding perl script here:
http://www.josupeit.com/weblog~informatik-und-technik~linux,spamassassin-anhand-von-imap-flags-lernen-lassen.html