Dave Pooser a écrit : >> manual training on any FPs/FNs that were not correctly autolearned from is a >> good idea. > > Also, it does no harm to re-learn messages that have already been learned, > so it's perfectly acceptable to set up a cron job to learn the contents of a > folder as spam and then just drop FNs-- or even accurately-flagged spam-- > into that folder for learning. You can do the same thing with ham, too. I > have my server regularly learn as ham a couple of users' inboxes to make > sure Bayes has a good ham corpus to work with.
while I have no problem training on: FNs, FPs and TPs, I found no "nice" way to train on TNs (True Negatives, that is, ham correctly detected as such). indeed I can see a user moving spam and FPs to specific folders (or doing some action), in general, one doesn't move TNs. one way would be to train all "hammy" mail (mail tagged as "ham" and not "corrected" by the user) as ham, and if it is spam, then "unlearn" it and retrain it as spam, but I don't know all the consequences of this. my guess is this is too risky/fragile.
