On Wed, 22 Jul 2015 09:52:10 -0400 Bill Cole wrote: > On 22 Jul 2015, at 8:18, RW wrote: > > > YMMV but personally I've never had a single ham hit BAYES_99. > > There's currently no evidence to suggest that the OP would have any > > problem with short-circuiting on it. > > Experiences with that absolutely do vary, widely.
That's rather my point. > Keep in mind that > Bayesian classification gives a statistical metric, not a human > claim; the delta from 100% isn't a polite warning, it's as hard a > fact as statistical prediction can provide, given a valid Bayes DB. > 99.00% spam certainty from Bayes will be wrong 1% of the time, on > average. This is at best a naive executive summary. None of the above is really true. > If you've actually NEVER had BAYES_99 hit on ham, you're > quite lucky or don't get a lot of ham. I get enough to know that for me the upper limit to the FP rate on BAYES_99 is negligible compared with the FP rate for SA as a whole. What's important is to compare the FP rate increase that would be caused by raising the score of BAYES_99 with the SA FP rate caused by the rule rescoring and custom rules that were added to avoid the FNs in spam that hit BAYES_99 . It's also useful to repeat that analysis without the FPs that no-one cares about. Unless you have done this you don't really know whether increasing the score of BAYES_99 is a good idea or not.
