When I frist read Paul Graham's plan for spam, I decided to write my own
Bayesian Spam Filter. The core engine was fairly easy to complete, and
it worked quote well.
I first wrapped it in a POP3 Proxy, and later, I turned it into a
Mailet. The Mailet was abit trickier because Iw anted to make sure each
user could have their own corpus of trained spam so that what one person
thinks of as spam wouldn't affect another.
But the biggest shortcoming, both of the POP3 Proxy, and the Mailet, was
how to handle feedback. I found that a Bayesian Spam Filter worksbest
when it could get access to the exact input for feedback (i.e. "this
e-mail was spam", "this e-mail was not spam"). My best solution was to
just forward the e-mail back using whatever e-mail client the user used.
But forwarding an e-mail altered the headers, sometimes in an
unpredictable way based on the e-mail client. I've not yet convinced
myself that it is a good solution.
Anyways, that's what I've got. I've not made it into any kind of a
SourceForge project yet because (a) I'm lazy, (b) there hasn't been much
interest and (c) there are plenty of client-side or
non-James-server-side solutions. Nonetheless, I'm willing to set up a
project for it and let others do with it as they please if there is
interest.
Regards,
Brian.