SA 3.0 Bayes reliability

[Andy Jezierski <aj...@stepan.com>]

Has anyone noticed a decrease in bayes ability to identify spam?  I
upgraded from 2.63 to 3.0pre2, copied the bayes files over and ran sa-learn
--import to upgrade the files to the new version.  Now that it's running in
production I'm seeing a lot of bayes_00 in spam messages, before the
majority was bayes_99.

Anyone have any thoughts?

Andy

Re: SA 3.0 Bayes reliability

[Daniel Quinlan <qu...@pathname.com>]

Andy Jezierski <aj...@stepan.com> writes:

> I hate to reply to myself but, I've noticed that the messages that have
> bayes_00 have a very strange bayes score.
> 
> debug: bayes: score = 5.55111512312578e-17

That's just a very small number expressed in scientific notation.

Try rescanning a formerly BAYES_99 message or two and see what they get
now that you've upgraded.  Saying "it seems like more have lower Bayes
scores now" is anecdotal, comparing apples to oranges, etc. so we can't
do anything with that.

Daniel

-- 
Daniel Quinlan
http://www.pathname.com/~quinlan/

Re: SA 3.0 Bayes reliability

[Andy Jezierski <aj...@stepan.com>]

Andy Jezierski <aj...@stepan.com> wrote on 07/21/2004 01:24:54 PM:

> Has anyone noticed a decrease in bayes ability to identify spam?  I
> upgraded from 2.63 to 3.0pre2, copied the bayes files over and ran
sa-learn
> --import to upgrade the files to the new version.  Now that it's running
in
> production I'm seeing a lot of bayes_00 in spam messages, before the
> majority was bayes_99.
>
> Anyone have any thoughts?
>
> Andy
>

I hate to reply to myself but, I've noticed that the messages that have
bayes_00 have a very strange bayes score.

debug: bayes: score = 5.55111512312578e-17
debug: bayes: score = 3.33066907387547e-16
debug: bayes: score = 0

All of the individual token tests seem to have good scores, it's just the
final score.

Anyone?

Andy