kalin mintchev wrote:
hi all...
using dspam_train on a qmial, vpopmail, dspam machine.
question #1:
which user should i use with dspam_train when training it for the first
(and later on too) time so that all virtual vpopmail users benefit?
Whatever user that you like. You will then use that user to seed a
group (check the README for groups), with something like
mergeduser:merged:*
Which will provide merged data, from mergeduser, to *.
question #2:
what would be a decent ratio spam:nospam to maintain when
training/relearning.
As close to real life as can be.
question #3:
apparently when sending mail to a relearning spam alias the user specified
on the 'command line' can be any virtual user. does that mean that after
the message sent to it is processed for relearning all the other users now
are also learning from this correction?
If *UIDInSignature is turned on, then the --user is changed on-the-fly,
yes. The originally specified --user has simply to exist, or else dspam
will bail (again, in the README, or in doc/something.txt)
the thing is like if the spam alias is used for a particular virtual user,
how would i know what is the total 'virtual users' ratio of spam:dspam so
i'd know that i have to feed another number of non spam to dspam to learn
and not watering down the accuracy of the filter...
I wouldn't worry about this. So long as you don't have a huge upset,
you should be OK. 100 spam and 50 ham would be OK, but 10,000 spam and
1,000 ham probably wouldn't. You would use dspam_stats to view your
current data.
thanks...
Mhm.
