I haven't found anything in the FAQ specifically about scoring
techniques for spam.
Seems to me that the single-factor tests which we're presently using
in the system filter sometimes cause false-positive rejections of
genuine mail that happens to hit one of those single-factor tests.
Most of the spams that would match these single factors will have
other indicators that they are spam.
On the other hand there are some single factors associated with spam
that are such weak indicators in isolation that we couldn't
practically use them on their own.
Consequently I was thinking it would be useful to accumulate some of
these indicators by a series of tests, maybe even both positive and
negative indicators of spam-likelihood (similar to what one would do
in a usenet scorefile, for example), and was a bit surprised to see no
mention in the FAQ.
A search threw up a 1998 answer from Phil on the potential usefulness
of user variables:
http://www.exim.org/pipermail/exim-users/Week-of-Mon-19980126/006454.html
(which, evidently, have meantime been implemented), and a personal
suggestion from a Matthew B-M:
http://colondot.net/mbm/mailfilter.shtml#score
and a few other only vaguely related hits. I was rather surprised not
to find more. Have others found the technique useful? Are there
hidden snags?
Perhaps it needs a neural net, after all ;-)
best regards
