Functional iterative approach for Universum-based primal twin bounded support vector machine to EEG classification (FUPTBSVM)

Abstract

Due to the increasing popularity of support vector machine (SVM) and the introduction of Universum, many variants of SVM along with Universum such as Universum support vector machine (USVM), Universum twin support vector machine (UTSVM), have been applied to several binary classifcation problems like electroencephalogram (EEG) signals, handwrit ten digit recognition and many more. Universum, which is not belonging to either class, is considered recently by many researchers that accept the prior knowledge into the binary classifcation process. In this paper, an efective and improved approach of TSVM with Uni versum data is proposed named a functional iterative approach for Universum-based primal twin bounded support vector machine to EEG classifcation (FUPTBSVM) which provides better performance. It also considers the gist of structural risk minimization (SRM) theory through the inclusion of the regularization parameter in the primal problem of FUPTBSVM and solved through a functional iterative approach. The regularization parameters terms are added to enhance the stability and make the model well-posed. Our proposed approach FUPTBSVM along with four standard classifcation approaches is tested on various EEG signals datasets with N Universum data or without N and benchmark real-world datasets. After conducting several numerical experiments with our proposed algorithm, one can ana lyze that FUPTBSVM improves the generalization performance in comparison to USVM, UTSVM, RUTSVM, and ULSTSVM for binary classifcation problems using Gaussian ker nel. It achieves 88.66% accuracy which is higher than other compared approaches for real world datasets. It is also computationally intensive among all concerned approaches.