Peter Marshall wrote:
I have been running spamassassin 3.0.2 for the last month or so. I have bayes set up to run for each user. It is working very well. However, I never get any missed marked spam (ie, messaged marked as SPAM that should not have been), I only get a few messages that don't get flagged. My problem is that I read you have to teach the bayes filter with both spam and ham. I have on our imap server for each user, a Missed Spam and a Not Spam folder. Each user is supposed to put messages that are spam but not flagged into the "Missed Spam" folder, and are supposed to take miss flagged messages and put them in the Not Spam folder. I have a script that runs to learn the messages. I was just wondering, is it ok, to just drag messages from my Inbox into "Not Spam" so that I learn some ham as well ? or do I need to do this.
There's no single school here. I personally only train on errors. In standard bayes, this is supposed to limit the problem of overfitting. now, SA isn't a "standard" bayesian filter. an alternative is to reset the bayes db and retrain with recent messages (both ham and spam). but that's more work than "train on error"...
now, a simpler rule is: if it works, don't (un)fix it.
