On Tue, 13 Feb 2007, LuKreme wrote:
Now, perhaps I am misunderstanding, but BAYES_99 is hitting on 5% of ham? and AWL on 35% of spam?
Keep in mind that AWL is slightly misnamed; it doesn't just whitelist, it adjusts scores (both positively and negatively) based on previous history. So the fact that it's hitting on 35% of your spam is pretty meaningless, really. sa-stats counts something as spam that SA marks as spam. So the fact that BAYES_99 is hitting on 5% of ham means (roughly) that 5% of your unmarked mail hit either only BAYES_99 or BAYES_99 and not enough other rules to mark it as spam. That means, respectively, that either you need to work on training your Bayes better, or that your Bayesian component is very well trained and that you need to turn up the scores for BAYES_99. The only way to know the difference is to look at the messages that are getting tagged with BAYES_99 but are not marked as spam. If Bayes is right about them, turn up your scoring; if not, continue training. This is where a user feedback look -- such as spam/ham reporting links in your webmail client, or the equivalent training for desktop client users -- can be really useful. Chris St. Pierre Unix Systems Administrator Nebraska Wesleyan University ---------------------------- Never send mail to [EMAIL PROTECTED]
