Error bounds of decision templates and support vector machines in decision fusion

Abstract

The need for accurate, robust, optimised classification systems has been driving information fusion methodology towards a state of early maturity throughout the last decade. Among its shortcomings we identify the lack of statistical foundation in many ad-hoc fusion methods and the lack of strong non-linear combiners with the capacity to partition complex decision spaces. In this work, we draw parallels between the well known decision templates (DT) fusion method and the nearest mean distance classifier in order to extract a useful formulation for the overall expected classification error. Additionally we evaluate DTs against a support vector machine (SVM) discriminant hyper-classifier, using two benchmark biomedical datasets. Beyond measuring performance statistics, we advocate the theoretical advantages of support vectors as multiple attractor points in a hyper-classifier's feature space.