Chris Santerre wrote: >> I thought I remembered a few years back that Baysian chains had a >> 10% increase in capture rate over straight Bayes rules. I would >> think that this is similar.
Marc Perkel wrote: > I've always thought that a second basian filter that would just > look at rule hits would be worth trying. No message content, just > rule combinations. SA uses "naive Bayes" which mathematicians sometimes call "idiot Bayes" for its simplicity. No priors are used, so everything is judged at the same level. That said, SA's point-based system actually accomplishes this to a degree, since rules are automatically re-scored based on their likelihood of matching ham. This is basically Bayesian, though simplified into a point-based system rather than a probabilistic one, presumably for computational simplicity. As to how much ground would be gained migrating from a point-based system to a probabilistic one (versus the added complexity), that's a great question. I'm sure ground would be gained, though I'm not sure if it would warrant the change, especially as it makes writing custom rules quite difficult (since they'd need to be thrown into the wild and automatically scored, despite that custom rules are often written to address short-term problems). I'm going to refer the the current bag-of-words bayes db as bayes-words and the SA rules-rescoring bayes db as bayes-scores. Now to really dig into the issue: Typically, Bayes uses users' evaluation of their own data to judge their own future data. SA's scoring system uses ~developers' evaluation of masscheck data to judge third-party future data. If we move to a Bayesian scoring system of the rules themselves, whose input does it reflect, and is that wise? My immediate thought is to use the best of everything. The official masscheck bayes-scores db could be pulled by each deployment regularly and merged (perhaps with some user-configurable weight favoring the masscheck db, defaulting to 70% masscheck) with the deployment's own bayes-scores db and the same for the bayes-words dbs (perhaps the user-configurable weight on words would be 30% masscheck). The bayes-scores result acts as a prior for the bayes-words test, and the final outcome is derived from that. Rather than merging masscheck probabilities with local probabilities, another option is to take four weighted passes.
