When h=f you would choose e(h,f)=0. When

you need to distinguish between the false positive and the false negative and administer an error (penalty) accordingly. That is where the risk matrix comes in. It tells you how to penalize the different errors.
Quote:
Originally Posted by LowEntropy
I guess I'm confused with how to apply the equation from the slides:
Ein = 1/N sum( e( h(x) , f(x) ) )
where e( h(x), f(x) ) == 1 or 0
To the problem with weighted cost values. Where do the weights from each possible choice fall into play in finding the in-sample error for the problem of classification? Thanks!
|