You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@spamassassin.apache.org by Gregorics Tamás <ta...@mmcomputer.hu> on 2007/06/18 08:22:28 UTC
Bayes giving false positives
Hi!
I have a problem with bayes' scoring. It gave BAYES_99=3.5 to a mail
which is not a spam. Unfortunately with this addition it reached my
required score so it got classified as spam.
How can i fix this behavior? Only auto learning is enabled with the
default threshold, no one could possibly feed it false data.
Thanks
Re: Bayes giving false positives
Posted by Kris Deugau <kd...@vianet.ca>.
Gregorics Tamás wrote:
> I have a problem with bayes' scoring. It gave BAYES_99=3.5 to a mail
> which is not a spam. Unfortunately with this addition it reached my
> required score so it got classified as spam.
>
> How can i fix this behavior?
Tweak the autolearn thresholds a little.
> Only auto learning is enabled with the
> default threshold,
This statement ...
> no one could possibly feed it false data.
... is in direct conflict with this one, IME. The default thresholds
*can* allow incorrect autolearning of very hammy spam, or spammy ham.
I'm not sure what the defaults are now, but I've run with 12 and -0.1
for quite a while with very little trouble - previously, the default
autolearn-as-ham threshold of 0.1 actually got a few very low-scoring
spams learned as ham. (around about SA2.55, IIRC) And I *have* seen
nominally legitimate email scoring in the 12-15 range on occasion. :(
-kgd
Re: Bayes giving false positives
Posted by Theo Van Dinter <fe...@apache.org>.
On Mon, Jun 18, 2007 at 08:22:28AM +0200, Gregorics Tamás wrote:
> I have a problem with bayes' scoring. It gave BAYES_99=3.5 to a mail
> which is not a spam. Unfortunately with this addition it reached my
> required score so it got classified as spam.
>
> How can i fix this behavior? Only auto learning is enabled with the
> default threshold, no one could possibly feed it false data.
sa-learn --ham
auto-learn does a good job, but "learn on error" is always recommended.
--
Randomly Selected Tagline:
lp1 on fire
(One of the more obfuscated kernel messages)