[EMAIL PROTECTED] (Justin Mason) writes:

> Would it be possible to have *both*?  Namely one set of rules (with
> high scores) for the -firsttrusted case, and another for the non -
> -firsttrusted case (with lower scores).  The latter are very good for
> catching spam via mailing lists, for example.

Ugh, what a can of worms.  I'm not sure there will be enough hits of the
latter type to get good scores from the perceptron.  The S/O ratio for
the latter type is also much lower once you subtract out the
-firsttrusted hits in a correctly configured setting (counting Michael
Parker as misconfigured, although I'm not sure whether it's a trusted
networks guessing problem that we should be able to guess or not, but
we're currently not guessing his trusted networks right).
 
Daniel

-- 
Daniel Quinlan
http://www.pathname.com/~quinlan/

Reply via email to