Michael Monnerie wrote:
> On Dienstag, 9. Mai 2006 23:14 Bowie Bailey wrote:
> > When I look at the overall stats, bayes does pretty good:
> > RANK    RULE NAME   COUNT %OFRULES %OFMAIL %OFSPAM  %OFHAM
> > ------------------------------------------------------------
> >    6    BAYES_99    26754     4.19   44.49   67.00    3.06
> 
> 3% HAM hits for BAYES_99 is horrible, not good. It's the FP that
> should make you alert.

True enough.  But no complaints so far.  I'm not sure how many of my
clients are even taking advantage of the spam markup.

> > But when I do it for only our domain (which is where all the manual
> > training happens), it hits less ham, but less spam as well:
> > RANK    RULE NAME   COUNT %OFRULES %OFMAIL %OFSPAM  %OFHAM
> > ------------------------------------------------------------
> >    8    BAYES_99     4649     3.29   33.41   54.64    0.20
> 
> At least much better FP rate, by a factor of 15!
> 
> > Just my personal email address (which is trained aggressively) gets
> > very few ham hits (partly because I lowered my threshold to 4.0),
> > but less spam than overall: RANK    RULE NAME   COUNT %OFRULES
> > %OFMAIL %OFSPAM  %OFHAM
> >    ------------------------------------------------------------ 5  
> > BAYES_99     1643     3.08   27.05   65.72    0.08 
> 
> Again the FPs reduced...

Of course, it's being constantly trained and the spam threshold is
lower.  I am curious why I don't get more spam hits with a
well-trained database.

> > And then when I modify sa-stats to exclude our domain, I find that
> > our customers (who are trained exclusively with autolearn) seem to
> > do better than us: RANK    RULE NAME   COUNT %OFRULES %OFMAIL
> > %OFSPAM  %OFHAM
> >    ------------------------------------------------------------ 6  
> > BAYES_99    22105     4.44   47.83   70.35    4.11 
> 
> No, 4% FPs is nothing you should be happy with.
> 
> > Based on these results, it almost seems like the more training Bayes
> > gets, the worse it does!
> 
> But remember that sa-stats can never tell if that HAM/SPAM are really
> such, it just tells you what it *believed* was HAM/SPAM.

Right. That's what I was referring to below.

> > Are these anomolies just an artifact of sa-stats relying on SA to
> > judge ham and spam properly?  Can these numbers be trusted at all if
> > my users don't reliably report false negatives and positives?
> 
> As I said on the other thread: Be very careful what you feed to bayes.
> Try to find those 4% of FPs, and if they are really FPs. Maybe your SA
> made the mistakes because you don't have enough rules to detect all
> SPAMs.

The group with 4% false positives is trained exclusively through
autolearn.  There is no facility for manual training with those
accounts.

If I follow the false positives, it lines up with expectations.  The
more manual training in the group, the lower the false positives.  Why
don't I see a similar trend with the spam hits?

-- 
Bowie

Reply via email to