Support vector machine classification using semi-parametric model

Abstract

Pattern recognition and data mining using support vector machine (SVM) have been the focus of widespread researches in recent decades. In SVM, a hyper-plane is designed to classify the training data. A challenge in SVM is that the parameters of hyper-planes are constants. As a result, there may be some critical points that will be classified into a wrong set. It should be mentioned that finding this hyper-plane is very similar to solving a regression problem using parametric or semi-parametric models in statistics. This is the main motivation of this paper. The contribution of this paper is combining SVM classifier and semi-parametric models (SP-SVM) to solve the aforementioned challenge. In fact, using semi-parametric linear model results in some serial linear decision boundaries with several slopes and intercepts. In other words, there are two types of kernels in the proposed SP-SVM: the kernels that perform nonlinear transformation of the input features and the kernels needed in the semi-parametric model. The validations have been done on Iris data set and also some other linearly non-separable classification problems. The accuracy of the proposed SP-SVM outperforms some related algorithms such as K-nearest neighbor (KNN)-based weighted multi-class twin support vector machines (KWMTSVM), support vector classification–regression machine for K-class classification (K-SVCR), twin multi-class classification support vector machines (twin-KSVC), intelligent particle swarm classifier (IPS-classifier) and random forest. The accuracy of SP-SVM is 97.33%. Thus, SP-SVM can play an important role in increasing the accuracy of industrial machines that perform classifications, for example, agricultural products.