
Re: EFD vs EWD in a Naive Bayes Classifier
Posted:
Sep 19, 2008 9:08 PM


On Sep 19, 8:41 pm, Peter <peterp...@hotmail.com> wrote: > I do not know this particular application, but consider that, to keep frequencies equal, highdensity events will be binned in smaller intervals. So probability is inversely related to the length of the interval (as in a2a1) that your observation p falls in.
Define pi = P(p in bin i), li = length of interval i, ni = instances in bin i, k = constant of proportionality Then
pi = k * ni / li sum(pi) = sum (k * ni / li) = 1 1 = k * sum (ni / li) k = 1 / sum (ni / li)
therefore pi = ni / (li * sum (ni / li))
I suppose this makes sense from a mathematical derivation point of view, I just find it hard to see the intuitive view of it.

