Clustering-Based Hybrid Approach for Multiclass Classification Using SVM

Abstract

To develop a decision support system, various soft computing techniques are used like evolutionary computation, machine learning, deep learning, etc. To develop effective decision support system, classification and clustering is used in various applications. SVM is a binary classifier technique as it is based on separating hyperplane. Over time to enhance the applicability of SVM in multiclass classification, some methods have been discovered, but results in multiclass classification through SVM are not impressive like binary classification. This paper presents a clustering-based multiclass SVM classification. In this approach, to obtain better accuracy in multiclass problem, k-means clustering is used to partition the datasets based on properties of conditional attributes and then the SVM multiclass classifiers are developed for each partition. To make decision about testing instance, nearest cluster in each instance is searched and then the corresponding trained SVM model is used to assign the class value of instance. The simulation results of proposed approach show superior accuracy rate for both one-versus-one and one-versus-all SVM methods in comparison to normal SVM-based multiclass classification.