Unknown class detection in the gulf classifier

Question

If you have a bay classifier trained for a set of classes, how do you determine if there is enough output to select a class? It would be useful to discover patterns that cannot be assigned to a class. I tried to test if the probability of the class is above average + 2 * stddev of the probabilities of all clans, but I do not think that it will be reliable.

Answer 1

You can use log-likelihood coefficients. Consider R(C) = log(P(C|D)/P(~C| D) , where C is a class, D are functions. Then you probably want to make sure that R(C) greater than some positive sum .