Michael Monnerie wrote:
> On Mittwoch, 10. Mai 2006 17:08 Bowie Bailey wrote:
> > Yes, this user is set with all the default options for Bayes
> > learning and a spam threshold of 5.0.  The entire Bayes database
> > was created via autolearn for this user.
> 
> Is that possible at all? I though that bayes to work you need 200 ham
> + 200 spam first.

Sure it is.  Bayes will autolearn messages right from the start.  It
just waits until it has seen 200 ham and 200 spam before it starts
contributing to the score.  There is nothing saying that you have to
manually learn the first group of messages.

On the other hand, since there is very little direct feedback from
that initial set of messages, you have to be careful that false
positives and negatives do not corrupt the database before you even
get started.

> > It seems to me that Bayes is highly sensitive to the types of ham
> > and spam that each user gets.  This user has a near perfect Bayes
> > database created with autolearn.  No false positives or negatives
> > and 95% of spam hit by BAYES_99.  My account, on the other hand,
> > has a few false positives and only a 66% spam hit rate despite
> > aggressive manual training.
> 
> I had on offlist discussion with somebody, we tried to compare our
> setup and results. I'll post this as a separate thread tonight or
> tomorrow, I've gotta go now.

Sounds interesting.

-- 
Bowie

Reply via email to