Development of novel fuzzy least square twin support vector machines for classification problems

Abstract

Support vector machines (SVMs) have gone through significant research and achieved outstanding results in various applications. However, the success of SVMs is relatively restricted when learning from imbalanced datasets in which one class samples significantly exceed the other class samples. On the other hand, SVMs are also sensitive to outliers and noise in the datasets. As a result, even if the current class imbalance learning (CIL) approaches might reduce SVMs sensitivity to class imbalance, they may still experience outlier and noise issues. To address the aforementioned issues, several variants of SVM have been proposed. Unlike SVM, which solves a large quadratic programming problem (QPP) to get the hyperplane, a few variants, such as least square twin support vector machine (LSTSVM) and energy-based least square twin support vector machine (ELS-TSVM), solve the system of linear equations and hence are computationally efficient than SVM. Moreover, robust energybased least square twin support vector machine (RELS-TSVM) implements the structural risk minimization (SRM) principle, making it more robust to noise. Modifying the data distribution or the classifier are frequent solutions to the class imbalance problem. The fuzzy theory has been successfully used with machine learning models to address their issues with noise and outliers in the data samples. This thesis proposes two novel models to deal with noise, outliers, and imbalanced data problems.