Development of robust support vector machine algorithms with biomedical applications

Abstract

The work presented in this thesis comprises robust and efficient machine learning (ML) models based on novel optimization approaches. The ML technique investigated very thoroughly in this work is support vector machine (SVM). SVM is a widely used supervised learning algorithm for classification as well as regression problems. It uses a kernel based approach for efficiently classifying the data. Since SVM based algorithms have been extensively used for classifying biomedical data, we applied most of the proposed SVM models to biomedical applications. Our survey identified some key problems in training SVMs, viz. (which are) class imbalance problem, data with noise, no knowledge about data distribution, unlabelled data, and proper feature selection. Also, these problems often occur with biomedical data as well. To resolve these issues, we proposed novel SVM based algorithms involv ing universum data and fuzzy logic. We presented the applications of these models for healthcare, such as automated diagnosis of diseases like brain disorders. Moreover, we focused on the issue of proper integration of machine learning algorithms with specific applications. We also performed a review of the various machine learning techniques used in detecting brain disorders. This resulted in the detection of key problems re lated to machine learning usage in the biomedical domain. One of the key issues is identifying features containing possible locations of brain regions responsible for the disease. To resolve this, we proposed a novel feature selection technique based on universum SVM in this thesis. To deal with the problem of class imbalance, we proposed two novel algorithms for classification. One of the algorithm is termed as a robust fuzzy least squares twin support vector machine for class imbalance learning (RFLSTSVM-CIL). The RFLSTSVM-CIL algorithm removes the class imbalance problem using a fuzzy logic based approach, which in turn helps to deal with noisy data as well. The second algorithm is proposed using a different approach involving universum data to use prior information about data distribution for class imbalance scenarios. This algorithm is termed as a reduced universum twin support vector machine for class imbalance learning (RUTSVM-CIL). We utilized universum learning for neurological disorders in this work. For epilepsy, we used electroencephalogram (EEG) recordings to propose a universum SVM based seizure detection technique. Moreover, for feature selection, we proposed a universum based feature elimination algorithm, termed as universum support vector machine based recursive feature elimination (USVM-RFE). We applied the proposed USVM RFE on Alzheimer’s disease (AD) using magnetic resonance imaging (MRI) data for classification. However, to deal with noisy datasets in universum learning, three fuzzy logic based universum algorithms are proposed in this thesis as: A fuzzy universum support vec tor machine (FUSVM), a fuzzy universum twin support vector machine (FUTSVM) based on information entropy, and a fuzzy universum least squares twin support vector machine (FULSTSVM). However, universum learning incurs additional computation time. Therefore, we presented some efficient universum based SVM algorithms. We proposed an efficient angle based universum least squares twin support vector machine (AULSTSVM) for pattern classification. AULSTSVM uses an angle based approach for universum learning. Also, two novel variants of universum based twin SVM algo rithms are proposed as: Improved universum twin support vector machine (IUTSVM), and universum least squares twin parametric-margin support vector machine (UL STPMSVM). Most of the above mentioned algorithms are applied on Alzheimer’s disease data for detection of disease. To explore the domain of unsupervised learning, we presented an SVM based algo rithm, termed as least squares projection twin support vector clustering (LSPTSVC), and applied on AD data. All the models proposed in thesis are compared with baseline algorithms to justify the advantages. The results of numerical experiments are compared using statistical significance tests. Keywords: Support vector machine, class imbalance, universum data, fuzzy mem bership, Alzheimer’s disease, epilepsy, MRI, EEG, twin support vector machine, clus tering