Extensions of the SVM Method to the Non-Linearly Separable Data

The main aim of the paper is to briefly investigate the most significant topics of the currently used methodologies of solving and implementing SVM-based classifier. Following a brief introductory part, the basics of linear SVM and non-linear SVM models are briefly exposed in the next two sections. The problem of soft margin SVM is exposed in the fourth section of the paper. The currently used methods for solving the resulted QP-problem require access to all labeled samples at once and a computation of an optimal solution is of complexity O(N2). Several ap-proaches have been proposed aiming to reduce the computation complexity, as the interior point (IP) methods, and the decomposition methods such as Sequential Minimal Optimization – SMO, as well as gradient-based methods to solving primal SVM problem. Several approaches based on genetic search in solving the more general problem of identifying the optimal type of kernel from pre-specified set of kernel types (linear, polynomial, RBF, Gaussian, Fourier, Bspline, Spline, Sigmoid) have been recently proposed. The fifth section of the paper is a brief survey on the most outstanding new techniques reported so far in this respect.