[PATCH] bogofilter-0.7.6-gl1

[Greg Louis]

On 20021029 (Tue) at 1221:44 -0500, Clint Adams wrote:
> > A patch is available via http://www.bgl.nu/~glouis/bogofilter that
> > removes Graham calculations from bogofilter-0.7.6 (which supports both
> > Graham's original method and some of Robinson's proposed improvements).
> 
> Even though I am using the Robinson, the Graham fares better in my tests
> of other people's spam corpora.

A major comparison is in progress.  I've seen situations where Graham
wins, and it seems (don't hold me to this but I'm getting an
impression) that when the training sets are big enough, either method
does quite well.  Unsurprisingly, one area where Robinson excels is
recognizing duplicates of spams it's been trained on.

> You could use GNU diff's -D to produce #ifdef'd sources which could
> facilitate a configure-time --enable-algorithms=robinson.

On my TODO list.  I'd like to make either or both calculation methods
optional at compile time (obviously you need one :) and probably the
whole -R thing as well (David's histogram code is better if you don't
have or want R).

-- 
| G r e g  L o u i s          | gpg public key:      |
|   http://www.bgl.nu/~glouis |   finger greg at bgl.nu |