On Wed, 26 Jul 2006, [EMAIL PROTECTED] yowled: > My impression is that the perceptron tries to cluster scores NEAR 5.0 > with as much spam as possible over 5.0 and as little ham as possible > over 5.0.
Well, it doesn't *try* to cluster, but since it'll keep tweaking until as many FPs and FNs as possible are eliminated, it'll tend to end up with a pair of `humps' near 5.0, and a slight reduction around 5.0 itself. > That's like mashing too boobs into one cup. Lift and > separate is the answer. It works for me so why not others, too. Your metaphors are worrying, this may be Too Much Information, but, yes, that is rather similar to what the accumulated mail-score graph looks like. I *had* been thinking of it as like a jelly mould, but doubtless now you've poisoned my mind and I won't be able to look at that graph again without getting, ahem, distracted. :/ -- `We're sysadmins. We deal with the inconceivable so often I can clearly see the need to define levels of inconceivability.' --- Rik Steenwinkel