better Bayesian bogofilter

[Greg Louis]

On 20030813 (Wed) at 0843:44 -0400, David Relson wrote:
> 
> Remember that Greg wants to test the effect of accurate ratios vs 
> inaccurate ratios.  If the test shows that accuracy doesn't matter, then 
> there's no need to implement the feature.  If the test shows it _does_ 
> matter, that's the time to figure out the best implementation.

I consider that question to have been answered by the experiment
reported in the paper (BTW it's on the web with some corrected typos at
http://www.bgl.nu/bogofilter/bayes.html).  What we need to know now is
how current the ratio needs to be; last 2 weeks, last month, last 3
months or what?  The ratio in the population is changing rapidly: 30%
spam in January, 60% in June for my workplace.  It's important to test
how sensitive the discrimination is to minor discrepancies, in order to
learn how best to trade convenience for currency.

-- 
| G r e g  L o u i s         | gpg public key: 0x400B1AA86D9E3E64 |
|  http://www.bgl.nu/~glouis |   (on my website or any keyserver) |
|  http://wecanstopspam.org in signatures helps fight junk email. |