From: Steve Manes (no email)
Date: Sun Sep 01 2002 - 18:57:23 EDT
At 05:11 PM 9/1/2002 -0500, Schmehl, Paul L wrote:
>I think you need to reread Paul's thesis. Rather than using fifteen
>keywords, his filter calculates the "spam value" of the fifteen *most
>significant* words found in the message. IOW, it look at *all* the
>strings in the message - header and body - assigns values to each and
>every string and then sums the values of the fifteen most significant
>words.
What makes this work is how the filter learns from the messages you've
marked/deleted as spam and the ones you haven't, building two dictionaries
of the most significant words in the two sets of messages. If the filter
sees that you always mark as spam those messages containing the word
"madam", there's a high probability that all incoming messages containing
that word are also spam. The trick is to develop enough sample points for
the filter to sum the weights of words in a message and return a score,
rather than do what a conventional spam filter does and toss anything that
contains the phrase "dear friend", which could have innocent uses too. You
would tighten up what the filter delivers to your mailbox by reducing the
gateway score, say from 95% to 85%.
It's an adaptive rather than blindly reactive filter that learns what you
like and don't like. I could see this being deployed across a large ISP
too. It's a pretty cool idea.
-
To unsubscribe, send mail to with content
(not subject): unsubscribe postfix-users
|
|
|