> 
> Tony Godshall wrote:
> 
> >Hi, Justin, all.
> >
> >I'm doing nearly the opposite:
> >
> >My upstream runs spamassasin.  I run a a non-naive bayesian
> >(crm114) myself.  The default config for crm114 is to strip spamassassin's
> >headers, but I've found that the SA headers give crm114 very very good hints,
> >and have acheived (to me, subjectively) amazing accuracy, not just for
> >spam vs nonspam but for spam vs 20 nonspam mailboxes, including troublesome
> >spammish lists like [EMAIL PROTECTED]
> >
> >It seems to me that the insertion of the results of the bevy of SA
> >tests improve the
> >results of bayesian learning significantly.

On Mon, 14 Feb 2005 12:36:52 -0800, Marc Perkel <[EMAIL PROTECTED]> wrote:
> I agree. I think that ultimately a second bayesian filter that was
> trained on only rule names could replace the SA scoring and become self
> scoring rules.

But why only on rule names?  Why not let the learning filter have the
rest of the data too?

Reply via email to