You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@spamassassin.apache.org by Paul Boven <p....@chello.nl> on 2005/03/07 12:39:23 UTC
Bayes and msg-ids
Hi everyone,
Could anyone shed some light for me on how and when (and especially why)
Bayes often generates its own Message-Ids when learning, instead of
using the one provided in the message? I have a lot of Message-Ids that
are '@sa-generated' in my Bayes database.
This also makes it a bit hard to check if a message was indeed correctly
learned, because the real Message-Id never makes it into bayes_seen.
This is the scenario I'm worried about:
1.) A spam (e.g. a stock-spam) goes trough our filter-machine and gets
falsely autolearned as ham. I've seen this happen quite a few times.
2.) If the recipient happens to be one of our exchange-users, they
forward it back (as attachement). The mail is stripped out of the
attachement and fed back into bayes.
If the Message-Id is not the same when the spam-email comes around the
second time it does get learned as spam, but never gets to correct the
wrong auto-learning. In theory it would mean you could never get a
Bayes-probability over 50% for that particular spam... which indeed
seems to happen for some of the stock-spams of late.
Regards, Paul Boven.
Re: Bayes and msg-ids
Posted by Robert Menschel <Ro...@Menschel.net>.
Hello Paul,
Monday, March 7, 2005, 3:39:23 AM, you wrote:
PB> Could anyone shed some light for me on how and when (and especially why)
PB> Bayes often generates its own Message-Ids when learning, instead of
PB> using the one provided in the message? I have a lot of Message-Ids that
PB> are '@sa-generated' in my Bayes database.
PB> This also makes it a bit hard to check if a message was indeed correctly
PB> learned, because the real Message-Id never makes it into bayes_seen.
It's my understanding that this happens when SA could not identify a
Message-id for the email.
Perhaps Bayes should not auto-learn messages without message ids? Is
this already part of the Bayes auto-learn algorithm?
Bob Menschel