On Fri, 9 Mar 2012 08:38:21 +0100 Matus UHLAR - fantomas wrote: > >> On 05.03.12 12:15, RW wrote: > >> >I don't like it. It relies on FPs being removed from the SPAM > >> >folder rather than spam being sent to a learn-spam folder. > > >On Wed, 7 Mar 2012 15:35:05 +0100 > >Matus UHLAR - fantomas wrote: > >> Pardon me, but: > >> > >> Usage for end users > >> > >> *move mail into SPAM folder to classify as spam > >> *move mail out of SPAM folder to classify as not spam > >> > >> isn't the former what you want? > > On 07.03.12 21:44, RW wrote: > >I'm more concerned about what happens to the mail that isn't moved. > > apparently nothing, because it is assumed to be correctly evaluated.
So are you saying that a legitimate mail that hits BAYES_99 and scores 4.9 isn't worth learning as ham because it's correctly evaluated. > > >I think positive training is better than supervised autolearning > > those above clearly indicate postive and negative trainin, or do you > have different informations? When I first looked at it, it retrained on errors, with DSPAM autotraining on everything. It probably does support train-on-error, but IMO it would be inappropriate to train Bayes that way. > >The scheme might work well for pure train-on-error, but that's not > >really practical on Spamassassin where the classification is > >distinct from the Bayes result. > > pardon? If you're going to train on error then train on the right error, not a rarer, correlated error. The FP/FN rate based on the SA classification isn't anywhere near high enough to train BAYES. If a user receives 10 legitimate mails a day and SA works at its target FP rate of 1 in 2500, it would take over 100 years for Bayes to even turn-on.
