Bayesian Learning and forward to as spam action

[ian douglas]

> How can we train MS about false positive and negative ? I understand
> now that I don't have to worry about the learning process of spam/not
> spam but how about False + and false -. I don't understand how he can
> learn the good way when he was wrong :-)

I set up an IMAP account for one of my hosting users who gets a lot of
false-negatives (spam that slips through the filters), and he moves a copy
of the spam to this IMAP account which copies it back to the server where I
can run "sa-learn" on the mailbox to re-learn that certain messages are
still spam.

I have messages from him that have BAYES_99 in the header showing a full
point spread from SpamAssassin but since my "high scoring" rules stipulate
to delete the message if the total score is over 7.1, my BAYES_99 score
isn't high enough, and some spammers are clever enough to score on very few
things other than BAYES_99 anymore.

If I thought that my Bayes was perfected enough to not *have*
false-positives, I'd set my high score set at my BAYES_99 value to delete
messages that Bayes feels is spam.

-id