Randomtrain question
Jake Di Toro
karrde at viluppo.net
Mon May 19 17:44:04 CEST 2003
On Mon, May 19, 2003 at 11:16:51AM -0400, David Relson wrote:
> At 11:02 AM 5/19/03, Jake Di Toro wrote:
>
>
> >I'll save you from repeating my previous observations on my run, but I
> >figured I'd try to get this question answered.
> >
> >Is randomtrain not capable of working with the tri-state filtering? I
> >would think that when asking for specific training that an unsure
> >would not be suficient to classify messages as passing the spam/ham
> >testing.
>
> Randomtrain knows whether the message is ham/spam and checks bogofilter's
> result. A mis-match calls for training.
>
> Since "unsure" matches neither ham nor spam, any message classified as
> "unsure" is (by definition) incorrectly classified.
>
> I think you _can_ use it with tri-state, but there's no value in doing that.
I'm not trying to end up with unsures.... that's my problem. I ran
randomtrain with a tristate setup so that it would classify the
unsures (since that's how I normally run bogofilter). When I ran the
mail through bogofilter to verify the classification, the ham folder
was a mixture of No/Unsure results, and the spam folder was a mixture
of Yes/Unsure results (see my message 'Problem with randomtrain' for
details).
So apparently randomtrain isn't reclassifing a message when the result
is unsure. Or worse.. somehow my config options are being ignored to
use the fisher algo, because when I run the classification w/o my
config file it defaults to duo-state and all results are classified
properly.
--
Till Later,
Jake <karrde at viluppo.net>
More information about the Bogofilter
mailing list