[Maia-users] Per-user or per-system training

[Pete Barnwell]

On Thursday 21 September 2006 16:28, you wrote:
> Thanks Pete and Ryan for the replies...
>
> This would be ideal, if everyone had a common view.  There are a lot of
> spams arriving that refer to stocks and companies' financial performance,
> etc.  All of our users would classify such unsolicited e-mail as spam, but
> the language is very similar to some e-mails from subscription services
> that some users sign up for.  I suppose they should just whitelist those
> domains.
>
> Another problem we have had so far is that some users appear to be alert
> enough to confirm their delivered and spam messages, as they should, but
> lazy enough to not actually LOOK at the proposed classifications.  Thus,
> some spam messages are "confirmed" as ham.  And some such users get a LOT
> of spam, so a large number of spam messages are confirmed by them as ham.
>
> Have you dealt with this problem before?
>
> Also, according to the SpamAssassin configuration file documentation:
>
> "By default, each user has their own in their ~/.spamassassin directory
> with mode 0700/0600. For system-wide SpamAssassin use, you may want to
> reduce disk space usage by sharing this across all users. However, Bayes
> appears to be more effective with individual user databases."
>
> http://spamassassin.apache.org/full/3.1.x/dist/doc/Mail_SpamAssassin_Conf.h
>tml (bayes_path setting under "Administrator Settings").
>
> How tough would it be to make Maia/Amavis-new invoke SpamAssassin with a
> different bayes database based on the To: address of a message?

Very, I think - the way amavisd is written it is designed to run as a single 
user, and Maia inherits this behaviour. I find that a large % of my users 
never ever bother to log in and have a look at what is happening, so I don't 
have to worry about most of them mis-confirming things. The other issue I 
would have is that 99% of my users don't have an account on the mail servers, 
so they don't have a '~' , so you'd have to use some sort of common 
spamassassin settings anyway, which put's you pretty much back where Maia 
is...

This is the big advantage of a single user system - those that do train it 
effectively support those who don't, otheriwse those that don't bother 
training would find the system less and less useful as time went on and their 
bayes db became more and more useless.

I advise people who subscribe to mailing lists to whitelist the mailing list's 
from address - that eliminates false positives from the list. (If it's the 
sort of list where there is any amount of SPAM in it, then I'd seriously 
wonder if it were beneficial to be in it!

Rgds

Pete