Classifier selection based on support vector technique

Abstract

This paper presents an ensemble approach consisting of global SVM and local SVM. Global SVM is estimated according to its decision confidence. Local SVM handles the query whose global decision is of low confidence. Local SVM is constructed over query's neighborhood, which is developed under the guidance of an informative metric. And its training is based on a query-based objective function. Global SVM helps to define the new metric. Some heuristics proposed to specify neighborhood size and hyper parameters. We present experimental evidence of classification performance improved by our schema over state of the arts on real datasets. © 2013 Springer-Verlag.