Re: [exim] exiscan, spamassassin, and per-domain bayes datab…

Góra strony
Delete this message
Reply to this message
Autor: Steve Lamb
Data:  
Dla: Exim Users List
Temat: Re: [exim] exiscan, spamassassin, and per-domain bayes database
Marc Sherman wrote:
> I'm not convinced.


    I'm not trying to convince anyone.  I am pointing out that thinking "Gee,
spam for everyone is pretty uniform" and ignoring that ham for everyone is not
uniform is to misunderstand how a bayesian classifier works.  Furthermore it
is possible to have a bayes classifier on a global setting for several people
and not see problems with it because, as my message showed, a specific case
has to come up.  That case is where one's person's interests (and thus what
they consider ham) coincides with the spam that another person (or all people
since spam is pretty universal) gets.  So which of these fields regularly has
spam which would use it's keywords:
Programmer
Web Designer
Audiologist
Landscape Architect


    To be honest I just can't remember the last time I got spam offering me
hot Programmer tips, how to get the most outta my stereos, whether astroturf
or natural was better.  The closest I can come with the above list is that
there are times I get spam for OEM software which includes Dreamweaver.  So if
your Web Designer uses Dreamweaver you might have one term in there that gets
wonky.


    Contrast that with my example given.  Dad loves his eBay, he bids in or
runs several auctions a week.  By the same token eBay/Palpal phishing spam are
quire frequent.  There is a direct match of an interest he has that most of my
other users do not which fouls up terms that could otherwise be used to ID
spam in a heartbeat for those other people.  If you don't have such a direct
match I doubt you'd see much of a problem with just 4-5 people.



-- 
         Steve C. Lamb         | I'm your priest, I'm your shrink, I'm your
       PGP Key: 8B6E99C5       | main connection to the switchboard of souls.
-------------------------------+---------------------------------------------