Miroslav Vancl writes:
> 
> Now I found I'm not first with this idea. I had found the same idea from Tom
> Allison - 2 Jun 30, 2007 even with similar subject "A different approach to
> scoring spamassassin hits".
> 
> I'm sorry

Don't be sorry!  It's well worth experimenting with.

The best way to do this is to come up with a prototype that demonstrates
good results on "real-world" mail corpora.

I've tried it in the past, but it (surprisingly) didn't produce great
results.

My theory: we have very few "hammy" rules.  This is because it's trivial
for spammers to impersonate them, so it's safer to start from an
assumption of hamminess and allow attributes of the mail push it in one
direction only -- towards spamminess.  That works well for us, but not
so good for the bayes-like approach.  One possible fix for this would be
to consider the fact that a rule did NOT hit to be a token, as well as
the opposite.

--j.

Reply via email to