You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@spamassassin.apache.org by Alexander Litvinov <la...@ac-sw.com> on 2004/01/08 06:05:03 UTC
New rule type suggestion
Is it possible to write automatic program/script that will use mass check
results and write meta rules that score is some other rules triggered, like
multiple MUA, multiple RBLs and other written by hands.
It is possible that spam triggers few low-score rules but combination of riles
is very spam-specific. Is it possible to write script for determinating such
groups of rules ? If yes, it seems to me good idea to write meta rules for
this.
Re: New rule type suggestion
Posted by Mike Laiosa <mi...@void.recurse.net>.
I've been thinking about this for quite a while. The hypothesis is
that the SA rules are not linearly independent.
SA essentially implements a simple neural network. From that point of
view, adding one (or more) hidden layers would probably improve its
effectiveness. The problem with simply adding hidden layers is that it
makes it unclear how individual rules affect the final score, and the
ability to tweak rule weights is important to a lot of us. I figure
something like this would work best if it showed up in SA like the
bayes stuff does: an alternate classifier whose score is used as an SA
rule. Of course since the input to the multi-layer perceptron would be
which SA rules matched, there may be strange side effects when mixing
that result in with the rules in SA's own linear combination, but
experimentation is needed to find out for sure.
I was hoping to do some experimentation with that idea last summer, but
I got distracted with one thing, and then another, until my brain had a
stack overflow. Maybe I'll get around to it next summer, if someone
doesn't beat me to it.
Alexander's idea of automatically finding the linear dependancies and
representing them with meta-rules may prove to be a less invasive
solution to the problem. Neural network techniques (specifically
methods of training multi-layer perceptrons) may be a way to generate
the groups of rules Alexander is suggesting.
Mike
On Wednesday, January 7, 2004, at 09:05 PM, Alexander Litvinov wrote:
> Is it possible to write automatic program/script that will use mass
> check
> results and write meta rules that score is some other rules triggered,
> like
> multiple MUA, multiple RBLs and other written by hands.
>
> It is possible that spam triggers few low-score rules but combination
> of riles
> is very spam-specific. Is it possible to write script for
> determinating such
> groups of rules ? If yes, it seems to me good idea to write meta rules
> for
> this.
>