shrinking css files for quicker training?

Hi, all,

I've been mostly just letting my crm114 mailfilter installation just "sit"
for a number of years and it just keeps on working, of course.

However, along the way I think I may have expanded my css files too much,
perhaps too concerned about packing density.  I judge this from the fact
that the learning seems entirely too sluggish compared to what it used to
be.  

I assume so much has been trained by now that new training has less than the
desired impact.  This may have worsened due to the number of years the css
files have been in use, and just how much spam has changed over that time.
I would like some of the old training to "fall off the end" but I don't have
the time required to just start over right now.

I see nothing in the documentation to suggest that using cssmerge to move
back into a smaller file would not work, and I'm wondering what to expect
and how to optimize this.

I would rather avoid being forced into a bout of heavy training.  So I'm
thinking about reducing the size of the css files in relatively small steps,
(maybe 10% at a time?) and not worrying about power of 2 boundaries,
assuming that will work.  Or maybe I should just try going by a factor of 2
all at once?

I am currently at 1048577 buckets, and my current packing densities are 0.83
non-spam and 0.92 spam.  Maybe when I'm all done with this process I can
increase the size of my css files again to return to better densities.  But
if I were to reduce my css files by a factor of 8 in this process, I'd guess
off-hand I'd be running with the spam css at something like .99 density.  I
guess I'd find out along the way if this was a problem.

I suppose I could shrink and immediately expand the css in order to weed out
some of the old stuff.  Maybe I could then do this for the spam side only,
leaving the non-spam "experience" fully in place since non-spam is perhaps
under-trained anyway.

A few key settings from mailfilter.cf:

    :decision_length: /6000/
    :clf: /osb unique microgroom/
    :thick_threshold: /1.8/

And I'm using a quite old mailfilter, looks like a July 2006 build.  Some of
the newer developments were breaking my methodologies so I kind of had to
stick with mailfilter.

I'd appreciate any thoughts.

Of course I can just try some things, and see what happens, but for some of
these considerations it may be better not to be making random guesses.

-Kurt



------------------------------------------------------------------------------
LogMeIn Rescue: Anywhere, Anytime Remote support for IT. Free Trial
Remotely access PCs and mobile devices and provide instant support
Improve your efficiency, and focus on delivering more value-add services
Discover what IT Professionals Know. Rescue delivers
http://p.sf.net/sfu/logmein_12329d2d