If the final aim is to approximate the conditional information of the class given the attributes instead of optimizing the accuracy, then a different measure should be used. If conditioned to the attributes, what we obtain is a precise probability for the probability of the class , then the usual score for each case is the logarithm of the likelihood score: , where is the true value of the class. This scored is added for the different cases in the test set. If instead of a single probability, , what we obtain is a credal set , then there is not an obvious way of generalizing this score. Here we propose to use the following one: Let the probability having maximum entropy in , then we use the logarithm of the likelihood score computed with ...

Start Free Trial

No credit card required