From: "Bowie Bailey" <[EMAIL PROTECTED]>
jdow wrote:
From: "Bowie Bailey" <[EMAIL PROTECTED]>
> Michael Monnerie wrote:
> > On Dienstag, 9. Mai 2006 16:18 Bowie Bailey wrote:
> > > I've got per-user Bayes and most of my users
> > > don't bother to train it.
> >
> > Another reason for site-wide bayes, I'd say.
>
> I've considered that, but it won't work in our setup. This box
> scans our internal email as well as all of our customer's email.
> Since we are in an entirely different line of business from our
> customers, what we consider to be ham and spam will be quite
> different from theirs. If I could train it on both sets, it might
> work, but I don't have access to any of their emails for training.
>
> Also, I really prefer a per-user bayes for our internal email
> since there are various accounts that get a specific type of ham
> and work very well with Bayes.
Importune on them to feed you as large a collection of ham and spam
as they can, once. Then turn on autolearn, cross your fingers, and
put on your flack jacket.
What flack jacket? I have Bayes turned on now and I never did any
manual training on most of the accounts. I just turned it on and let
autolearn (with the default settings) do it's thing. So far, I have
received very few complaints.
But then again, I think less than half of my users are even taking
advantage of the spam markup. Since I don't do any blocking or
sorting on the server, it is up to them to use MUA rules to sort or
delete the spam once my server has marked it.
Fairly frequently I see evidence that autolearn can massively misfire
on SpamAssassin startup. It does not always happen or there'd be a lot
more messages about it. But there is apparently a vulnerable period
that can go bad with just the wrong selection of messages. Once the
database is large inertia will save the day.
{^_^}
