tri-state transition [was: bogofilter-0.93.0 - Significant Changes]

[David Relson]

On Sun, 07 Nov 2004 12:33:50 +0000
Robin Bowes wrote:

> David Relson wrote:

...[snip]...

> I've just upgraded - looks like I was the first to download from 
> sourceforge!
> 
> In the end I went with the new tags and modified my maildrop script to
> 
> look for "X-Bogosity: Spam".
> 
> I'm not sure that the "Unsure" status adds much to my situation. Let
> me discuss...
> 
> I run bogofilter in pass-through mode (-p) in a maildrop script, 
> filtering on the X-Bogosity header line. Spam (X-Bogosity: Yes|Spam)
> is put in the users' SPAM folder. Everything else (i.e. X-Bogosity: 
> No|Ham|Unsure) is delivered as normal.
> 
> I also use the -u option so messages are added to the wordlist 
> automatically.
> 
> If any spam is missed, the user can drop the message in a 
> SPAM/Undetected folder from where it is re-processed from a script run
> 
> from cron which essentially runs the message through bogofilter with
> the -Ns options.
> 
> Any messages mistakenly classified as Spam can be dropped in a 
> Spam/Misdetected folder from where they are re-processed using 
> bogofilter with the -Sn options.
> 
> So, I'm not sure how the "Unsure" classification contributes to my 
> configuration. I may even consider reverting to the binary
> classification.
> 
> Any suggestions?
> 
> R.
> -- 
> http://robinbowes.com

Hi Robin,

More good questions!  

I used the "-p -u" combination for a year or so. I noticed two things:
(1) my wordlist kept on growing, growing, ... and (2) many messages
scored as 0.000000 or 1.000000.

A partial solution for (1) was to control the size (somewhat) by
periodically compact the database (using bogoutil dump/load), though a
better solution would be to slow down the growth rate.

I theorized that (2) is a set of messages that are so clearly ham (or
spam) that little additional information is gained by training
bogofilter with them.  I added a "thresh_update=xxx" option to set an
update threshold so scores near 0 or 1 wouldn't auto-update and have
been running with "thresh_update=0.01" since early this year.  

One possible result is that scoring numbers differ more from 0 and 1,
i.e. messages start to score as 0.02, rather than 0.00.  When/if this
happens, bogofilter auto-corrects by training with the 0.02 message. The
net effect is that accuracy remains unchanged.

Using thresh_update has reduced wordlist.db's growth rate, as desired. 
It does not seem to have lessened accuracy -- for months now bogofilter
has been catching all spam, except for about 1 in 700.

Why do I mention all this?  Using tri-state classification lets me see
which messages bogofilter couldn't classify with certainty and lets me
manually classify them and be sure that bogofilter is properly trained
with them.

HTH,

David