Re: AW: After server reboot spam recognition rate wents down for some time

[John Hardin <jh...@impsec.org>]

On Mon, 2008-07-07 at 07:58 +0000, Tobias Eichner wrote:

> >From John Hardin:
> > Wild-ass guess: do you have autolearn enabled, and a reboot script that
> > clears the Bayes database?
> 
> Yes, autolearn is enabled, but I just did the standard installation via CPAN... not sure if this included such a script.

No, a script like that would *definitely* be custom.

One way to check would be to run "sa-learn --dump magic" after rebooting
and see if your nspam and nham counts are unexpectedly small.

But again, this was just a wild-ass guess, and likely isn't the problem.

> Anyway, I don't spend much attention to Bayes (bayes_99 increases score
> only by 1, since I had some bad experiences made with the reliability
> of Bayes on our system).

Poor bayes reliability is directly attributable to poor training. Bayes
works well if you leave the scores at their defaults and train properly.
I also always suggest that you retain your training corpa so that
retraining from scratch is possible, and reviewing mistraining is
possible.

If you have problems with bayes misclassifying, fix your database rather
than "fixing" the scores. If you're having problems with automatic
training, then move the scores at which messages are autolearned out a
bit (i.e. autolearn ham at a score lower than the default, and ham at a
score higher than the default).


-- 
 John Hardin KA7OHZ                    http://www.impsec.org/~jhardin/
 jhardin@impsec.org    FALaholic #11174     pgpk -a jhardin@impsec.org
 key: 0xB8732E79 -- 2D8C 34F4 6411 F507 136C  AF76 D822 E6E6 B873 2E79
-----------------------------------------------------------------------
  Homeland Security: Specializing in Tactical Band-aids for Strategic
  Problems.                       -- Eric K. in Bruce Schneier's blog
-----------------------------------------------------------------------
 Today: Robert Heinlein's 101st birthday