Partial classification: the benefit of indecision

Abstract

Classification methods may be improved in the sense of a meaningful, economically motivated benefit function, by allowing for indecision in a certain domains near the separation surfaces between the classes. Such a "partial" classifier, based on the intersection surface between parameterized probability density functions, is proposed. It is found to be beneficial with respect to "full" classification, assigning each new object to a class, in the prediction of stock behaviour.