Tom Meunier wrote:
> I'd say 98.4% accuracy is pretty darned good, Robin.
I agree - this is excellent! With version 1.42 I was getting around 85%.
> hold off a little, take the missed spam and send it through sa-learn.
I don't think that adding four messages to what the Bayes system has
learnt from will help much, but perhaps I am wrong.
> I don't know where you got the corpus of spam and ham from,
1832 non-spam emails, totalling 14.6 Megs, from the past 6 months.
1883 spams totalling 19,1 Megs, some going back to 6 months, but more
from recent times. These were partly those 85% found by the old
version, and a greater proportion than 15% of those spams which were not
found, so it was tweaked to optimise detection of messages version 1.42
missed.
These are for my own account - this is not a multi-account thing.
> but as you send current traffic through sa-learn, rather than a
> slightly stale corpus, or a corpus from another user, it'll get more
> and more accurate.
I agree, but I only have 4 spams to add to 1883 so far. Three of them
were rated pretty highly by Bayes already - and all of them were rated
significantly higher than any non-spams so far this week. So it seems
the best approach is to give Bayes postives more scoring power towards
the final result.
By experiment I found that tests, such as:
BAYES_40, BAYES_44, BAYES_50, and BAYES_56
(See: http://spamassassin.org/tests.html .)
which are set to score 0.001 will not show up in the headers. To make
them show up, but still not affect the score very much, make it 0.002.
I increased my Bayes scores with these lines:
score BAYES_40 -1.0
score BAYES_44 -0.5
score BAYES_50 0.5
score BAYES_56 2.0
score BAYES_60 3.5
score BAYES_70 5.0
score BAYES_80 6.0
score BAYES_90 6.0
score BAYES_99 6.0
and ran the original four undetected messages through SpamAssassin
again, from the command line. The extra Bayes scoring made three of
them become detected - over 5.0. As far as I know, I am not anywhere
near making good emails detect false positive, but that will take time
to determine for sure. So with this set Bayes scoring, I would have had
one false negative in a week of 260 messages, and no false postives!
I have documented all this at:
http://www.firstpr.com.au/web-mail/Postfix-SA-Anomy-Maildrop/
- Robin
-------------------------------------------------------
This SF.net email is sponsored by: Etnus, makers of TotalView, The best
thread debugger on the planet. Designed with thread debugging features
you've never dreamed of, try TotalView 6 free at www.etnus.com.
_______________________________________________
Spamassassin-talk mailing list
[EMAIL PROTECTED]
https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
