Hello all,

I am trying to come up with a good way to leverage one of our users'
extensive knowledge bases but I am not quite sure what the right
approach is.  All our users share similar mail behavior, but our
"power user" gets much more mail traffic than the others.

Currently, I have put everybody in a 'merged group' and our global
user was trained with dspam_train from some real corpus.  Each user
is set to TOE.  Accuracy was great for our "power" user, but not
quite as good for the other users since our global user does not have
that many tokens.

I thought it would be a good idea to dspam_merge "power user" to
"global user" to improve everybody's accuracy, so I did.  However,
accuracy -drastically- went down the drain.  Very obvious spam is now
tagged as innocent with confidence of over 95%.  The accuracy rate
went from over 99.5% to about 20%.

After reviewing the docs, I came to the conclusion that a
"Classification Group" would be a good solution: I'd just have to put
all our users in the same group, and they would all benefit from the
other user's data.

However, the documentation about classification groups is a little
confusing.  This is the behavior I would like (from the docs):

 "Classification groups allow a group of users to network their
results together.  If DSPAM is uncertain of whether a message is spam
or nonspam for a group member, all other members of the group are
queried.  If another  member believes the message to be spam, it will
be marked as spam."

Classification, shared and global groups seem (to me) to be confused
with each other in the docs.

--> My question is, how do I configure a classification group in the
"groups" file?   According to the docs, the "classification" keyword
is used for Global Groups (ie: groupname:classification:*globaluser), but
there's no mention of how to configure a "classification group"
as I want it (ie: users network their results together)

--> I don't want to use a Shared Group because of all the
implications it can have if people mishandle some messages.

Thank you very much for your consideration.

Tony

Reply via email to