Rob Winters wrote:

RW> SA does not give any credit to the cumulative effect that would be obvious
RW> to any human reading the "tests=" line, let alone the message itself. I
RW> mean, look at *this* one!"

This is currently true, and is basically a function of how the score-setting
(and score evaluation) works.  I have been doing some minor dabbling with
modified scoring schemes, where the combination of scores is not a simple sum,
but haven't spent long enough on it yet to really say anything definite about
whether there's a better way of scoring.  I'm sure there's some interesting work
here for any CS master's thesis if someone's interested.  Some kind of nonlinear
combination of scores maybe?  Or scores based not only on single rules, but also
on rule pairs, at least for the common pairings?

RW> X-Spam-Status: No, hits=4.0 required=5.0
RW> 
tests=SUBJ_ENDS_IN_Q_MARK,COPYRIGHT_CLAIMED,CLICK_BELOW,WEB_BUGS,MAILTO_LINK,RCVD_IN_OSIRUSOFT_COM,X_OSIRU_SPAM_SRC
RW> version=2.20
RW>
RW> It's freakin' toxic, and it gets a 4? SEVEN rule hits? Hmm... might be SPAM.
RW>
RW> How about a "bonus" for cumulative effect? Why not do a second-level
RW> analysis after scoring; something like:
RW>
RW> 3 positive score matches - add 1.0
RW> 4 positive score matches - add 2.0
RW> 5 positive score matches - add 4.0
RW> 6 positive score matches - add 8.0

This is I think a little too simplistic -- and if it were correct to do this,
it'd be reflected to some extent in the current scores.  It's slightly more
complex than the current scoring system, but I think one of the other
suggestions above would yield better results.  In particular, I think the rules
you mention above should take into account negative-scoring rules.  3 negative
scoring rules being triggered should yield a negative bonus, not a positive
one...

C


_______________________________________________________________

Don't miss the 2002 Sprint PCS Application Developer's Conference
August 25-28 in Las Vegas -- http://devcon.sprintpcs.com/adp/index.cfm

_______________________________________________
Spamassassin-talk mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/spamassassin-talk

Reply via email to