NEUROSVM: An Architecture to Reduce the Effect of the Choice of Kernel on the Performance of SVM

摘要

In this paper we propose a new multilayer classifier architecture. Theproposed hybrid architecture has two cascaded modules: featureextraction module and classification module. In the feature extractionmodule we use the multilayered perceptron (MLP) neural networks,although other tools such as radial basis function (RBF) networks canbe used. In the classification module we use support vector machines(SVMs)---here also other tool such as MLP or RBF can be used. Thefeature extraction module has several sub-modules each of which isexpected to extract features capturing the discriminatingcharacteristics of different areas of the input space. Theclassification module classifies the data based on the extractedfeatures. The resultant architecture with MLP in feature extractionmodule and SVM in classification module is called NEUROSVM. TheNEUROSVM is tested on twelve benchmark data sets and the performanceof the NEUROSVM is found to be better than both MLP and SVM. We alsocompare the performance of proposed architecture with that of twoensemble methods: majority voting and averaging. Here also theNEUROSVM is found to perform better than these two ensemblemethods. Further we explore the use of MLP and RBF in theclassification module of the proposed architecture. The mostattractive feature of NEUROSVM is that it practically eliminates thesevere dependency of SVM on the choice of kernel. This has beenverified with respect to both linear and non-linear kernels. We havealso demonstrated that for the feature extraction module, the fulltraining of MLPs is not needed. color="gray">