Unified SVM algorithm based on LS-DC loss

Over the past two decades, support vector machine (SVM) has become a popular supervised learning model, and plenty of distinct algorithms are designed separately based on different KKT conditions SVM model for classification/regression with losses, including convex loss or nonconvex loss. In this paper, we propose an algorithm that can train models in \emph{unified} scheme. First, introduce definition \emph{LS-DC} (\textbf{l}east \textbf{s}quares type \textbf{d}ifference \textbf{c}onvex) show most commonly used losses community LS-DC be approximated by Based DCA (difference algorithm), then unified algorithm, called \emph{UniSVM}, which solve any loss, only is computed, especially specifically chosen Particularly, training robust UniSVM dominant advantage over all existing because it closed-form solution per iteration, while always need to L1SVM/L2SVM iteration. Furthermore, low-rank approximation kernel matrix, large-scale nonlinear problems efficiency. To verify efficacy feasibility proposed perform many experiments some small artificial large benchmark tasks with/without outliers classification regression comparison state-of-the-art algorithms. The experimental results demonstrate achieve comparable performance less time. foremost its core code Matlab than 10 lines; hence, easily grasped users researchers.