On Fri, 2004-07-16 at 11:27 +0100, Tony Finch wrote:
> I note that on the page
> http://slett.net/spam-filtering-for-mx/x519.html
> you say: "Bayesian Filters. These require training before they are useful"
>
> This is not the case: SpamAssassin in auto-learn mode and without
> additional training is better than plain SpamAssassin, so is suitable for
> an unsupervised site-wide installation without per-user configuration.
it didn't take long before the site-wide Bayesian database was seriously
detrimental to SpamAssassin's performance here at the University of
Oslo. a site-wide database will only work if your user base is
reasonably homogenous, I think. we have students and professors from
all over the world, discussing all kinds of subjects in all kinds of
languages, and the database ended up giving most spam -4.90 points.
(this was around New Year, things may have improved.)
> I thoroughly recommend the following paper for a very thorough analysis of
> the leading open-source anti-spam software.
> http://plg.uwaterloo.ca/~gvcormac/spamcormack1.pdf
thank you for the link!
--
Kjetil T.
