On Jul 9, 2003 01:47 am, Kelsey Cummings wrote: > On Wed, Jul 09, 2003 at 10:24:55AM +1200, Simon Byrnand wrote: > ... > > > Certainly my initial impression after 24 hours is that DCC is the > > most effective, followed by RAZOR2, followed by PYZOR. All > > together is probably most effective of course, so long as the > > false positives of all three dont accumulate enough to register > > as spam. > > I've got a similar impression with my corpuses here, the DCC hit > rate appears higher than razor. One big advantage is you can run > your own DCC server and don't have to rely on the wonderful folks > who do an excellent job keeping the razor servers up (most of the > time.)
Looking at all spams received here recently (1103 total), I found the following hit totals for Razor2 and DCC (I don't use Pyzor):
DCC: 677 (61.4%) Razor2: 563 (51.0%)
The various RBL checks ranked as follows:
NJABL: 742 (67.3%) DSBL: 491 (44.5%) OSIRUSOFT: 446 (40.4%) OPM: 280 (25.4%) RFCI: 158 (14.3%)
And, for the Bayes scores:
BAYES_99: 2.4% BAYES_90: 71.5% BAYES_80: 14.3% BAYES_70: 7.9% BAYES_60: 3.9% BAYES_50 or less: none
Bayes has been trained with 5200/6200 ham/spams.
During the period that the 1103 spams accumulated, there was 1 FN and 0 FPs. This is great performance.
To the SA developers, I say: bravo!
Indeed.
However, are you able to generate any stats for the above tests (from the same period of time) for how often they triggered in *ham* ? EG so we can get an idea of the false positive rates of DCC, razor, and the various RBL's....sometimes the false positive rate is more important than the false negative rate. (Which effectively what your stats are displaying)
Regards, Simon
------------------------------------------------------- This SF.Net email sponsored by: Parasoft Error proof Web apps, automate testing & more. Download & eval WebKing and get a free book. www.parasoft.com/bulletproofapps1 _______________________________________________ Spamassassin-talk mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/spamassassin-talk
