RotaSVM: A New Ensemble Classifier

Abstract

In this paper, an ensemble classifier, namely RotaSVM, is proposed that uses recently developed rotational feature selection approach and Support Vector Machine classifier cohesively. The RotaSVM generates the number of predefined outputs of Support Vector Machines. For each Support Vector Machine, the training data is generated by splitting the feature set randomly into S subsets. Subsequently, principal component analysis is used for each subset to create new feature sets and all the principal components are retained to preserve the variability information in the training data. Thereafter, such features are used to train a Support Vector Machine. During the testing phase of RotaSVM, first the rotation specific Support Vector Machines are used to test and then average posterior probability is computed to classify sample data. The effectiveness of the RotaSVM is demonstrated quantitatively by comparing it with other widely used ensemble based classifiers such as Bagging, AdaBoost, MultiBoost and Rotation Forest for 10 real-life data sets. Finally, a statistical test has been conducted to establish the superiority of the result produced by proposed RotaSVM.