Some approaches I've taken:
* Set the autolearn thresholds low enough that most messages would be
learned. Not as accurate as human learning, but requires little effort.
* Set the minimum and maximum messages learned settings down to about 50
from their default 200. Not as accurate, but requires fewer messages in
the corpus to start scoring.
* Use server-wide Bayes and exclusively rely on autolearning. Again, not
as accurate but less work.
* Write a cron job to run sa-learn as the user, learning their inbox as
ham and spam folder (if they have one) as ham. Makes a leap of faith
that the inbox and spam folders will contain only ham and spam,
respectively, but again, requires little effort.
