Mailing List Archive

At 02:42 PM 3/11/2004, little@cs.ucsd.edu wrote:
>What exactly is the negative effect of having a ham corpus that
>is much larger than a spam corpus? Say we initialize both corpii (?)
>with equal amounts of spam and ham. However, with the majority
>of email being non-spam (or at least non-super-high-scoring spam),
>the auto-learn over time will cause the ham corpus to grow much
>larger than the spam corpus.

It's a complete misconception that your spam and ham training should be of
equal size.

Think about it.. bayes is a statistical probability system. It works best
when it's model is as close to reality as possible.

Therefore, I propose that the theoretical best bayes training ratio is not
1:1, but instead whatever your real-world spam:ham ratio is.

But frankly, bayes is pretty resistant to some considerable deviation from
this idea.

>The auto-learn feature is very attractive (us being, like most
>of you all, over busy and reluctant to take on more regular
>maintenance tasks), but is the only real answer to stay on
>top of spam, save it, and feed it regularly to sa-learn? It
>could be automated, but still... what's the point
>of auto-learn if it creates a less useful corpus set?

Autolearn isn't a replacement for manual training.. you can never have a
successful bayes database if you initialize manually and then rely
exclusively on autolearning. Autolearning doesn't do a good job as a sole
source of training because it never learns "mid-scoring" emails. And
because autolearning isn't bayes-based, your previous training doesn't help
the autolearner (this is to prevent "runaway bayes self-feeding" problems)

However, autolearning is a great supplement to your regular manual
training, and reduces the frequency with which hand feeding is needed.

Personally, I handle mine with the following:
autolearning on, but thresholds widened slightly.
email sent to select spamtraps is archived into a separate mbox
email sent to select nonspamtraps is archived into a separate mbox
both the spam and non-spam mboxes are fed to sa-learn with a daily
cron job, and then rotated into a multi-day storage archive.
the subject lines of the spam and nonspam training are emailed to
me for review. Any mistakes are corrected by hand.