Jeff Anderson wrote:
Hi Tom,
I don't post very often to the list, since there are others around that have
more expertise and time than I do to offer suggestions and fixes. I also
haven't used any of what I learned in my university statistics courses for at
least 10 years :) But I thought I might weigh in on this idea of multiple
statistical filters to try and achieve better accuracy.
While I like the idea of a multi-layered approach, we do so using some of the
built-in features of Postfix, then Amavis, and then DSPAM. 90% of our site's
(175 users) spam is blocked before it even reaches DSPAM. It works pretty
well and we're way over 99% accuracy.
After reading through the reasonable posts on the matter...
Not sure what you mean here by blocked. I'll assume you mean "never delivered
and never quarantined", so it's effectively lost email... Some operate in that
manner, others do not. It's a matter of opinion.
Working under the assumption that you do not entire discard email...
It seems to me that the best you can do is leverage multiple approaches
(amavis/spamassassin, bogofilter, spambayes...) into adding a token (X-HEADER)
to the email that is passed into the last dspam instance (the one that does the
quarantine) to add additional statistical weight to the emails score by
providing at least one more token with a very high contribution
(spam_messages/ham_messages is either ~1 or ~0) towards the end result.
Eventually I think I will work towards this end.
