On Tue, Nov 11, 2003 at 09:59:41AM -0800, Edward Craig wrote:
> I get the impression that when I forward a message (headers
>expanded) to [EMAIL PROTECTED] it sets my Baysian filter up, and anything it
>sees that looks like this message gets a spam rating based on its
>similarity to the forwarded message. Do repeated messages which fail the
>test add to the negative effect on subsequent similar messages?
>
At the moment spam samples forwarded to [EMAIL PROTECTED] are individually
examined by the BOFH to confirm that they are sample mails and not,
e.g., tech support questions (this test revolves around the issue of
whether they address content directly to an admin). Sample mails
are placed into a multi-gigabyte directory of sample spams i've
received since 2001. I need to feed them to the Bayesian learning
mechanism, but have not done so yet, as we've only recently turned
on Bayesian support and it's bound to be a lengthy process.
> Now, because efn users are way too human, they're fully capable of
>mistaking somebody's legitimate email with a "Subj: Hi" with spam, sight
>unseen, for some reason (like maybe seeing way too many spams with a
>"Subj: Hi" but instead of an introduction to someone or something of
>interest, an illustrated discussion of attribute enhancement. Graphic
>illustration, abbreviated discussion.).
>
> So the opposite of spam is tofu, thus the counter-spell to a
>false spam is to send it, headers expanded, to [EMAIL PROTECTED] This should
>take the bad luck off, but you may need to repeat, and repeat again and
>again as necessary.
>
That is correct; at the moment they build up in my tofu queue,
which will also have to be fed to SpamAssassin.
> Given the amount of spam coming at us, and how automated I can get
>about forwarding spam, I'm surprised I notice as few false positives as I
>do for spam from SpamAssasin and what I'm beginning to regard as my
>Baysian filter. My conjecture is a similar process for you, so forward the
>Dean stuff to [EMAIL PROTECTED], as you would spam to [EMAIL PROTECTED]
>
Yes, if you are getting incorrectly-flagged messages (e.g. from the
Dean campaign) please send them to [EMAIL PROTECTED]
On another note, in response to a point Marc raised earlier, we've
turned down the maximum score that the Bayesian component can give
to a potential spam; Bayesian results are no longer sufficient to
qualify a message as spam without other hits.
--
Networks are like sewers: my job is to make sure your data goes away when
you flush, and to stop the rats climbing into your toilet through the pipes.
-- Tanuki describes network administration
_______________________________________________
EuG-LUG mailing list
[EMAIL PROTECTED]
http://mailman.efn.org/cgi-bin/listinfo/eug-lug
