Classifiers with Low Decision-Making Error using Linear Combination of Functions.
Gérald Gavin, Didier Puzenat, Djamel Abdelkader Zighed · 1999
Usually in classication, the denition of the \\error rate" does not dierentiate an element misclassi ed from an element not classied. However, in some applications as medical diagnosis, it is better not to classify rather than to make a mistake. In such a case, a human can classify the element non classied by the learning system, eventually after further investigations (e.g. in the medical case, a deeper evaluation of patient history). In this paper, we will dene the decision-making error as the conditional probability that an element is misclassied knowing it is classied. We propose an algorithm, based on convex linear combination of classiers, in order to improve the decision-making error without increasing too much the not classied rate. We derive theoretical statistical results, on condence bounds for the generalization performance of linear combination of functions, to give condence bounds for our algorithm. Keywords: Classication, linear combination of functions, VC...