snowcrash+spamassassin wrote: > i noted in a recent thread a suggestion to not feed bayes-poisoning > spam to sa-learn. I missed that thread, but IMNSHO, that's Horsehockey.
Please, ask them to explain how well "bayes-poisoning spam" works in a system with chi-squared combining? Actually, ask them first if they understand what chi-squared combining is. > > that's an interesting thought; and actually makes some initial sense > to me. > > is this, in fact, widely suggested/recommended? I would highly advise against performing any "cranial-rectal-insertion" maneuvers that involve poisoning your own bayes database by feeding it unrealistic information. Train it spam as spam, ham as ham. Period. Anything else is just poisoning your own bayes DB. Think about it, if you decide a message must be "bayes poison", and therefore skip training it, the bayes poison has worked. Your bayes DB is now less prone to mark such emails as spam than it should be. And all of this, not because of the "poison" effect, but because of your actions.