Large Margin Cellular Piecewice Linear Classifier

piecewise linear classifiers have attracted a lot of attention in recent years, because of their simplicity and classification capability. In this paper, a large margin cellular piecewise linear classifier is introduced, called Cell-SVM. The cellular structure of Cell-SVM obtains a piecewise linear decision boundary which handles non-linearly separable data. Unlike the conventional SVM approaches, the proposed method employs multi hyperplanes instead of one in search space and resulting cellular structure addresses some important issues in machine learning such as: multi-modal classes, nonlinear classification, noisy data and outliers, small sample size, multi-class classification and overfitting to training samples. In experiments, we demonstrate significant gains for the well-known benchmark real datasets when compared to the usual multi-class SVM techniques with RBF kernel like OvO SVM, OvA SVM and MC-SVM. Besides, it is shown that the proposed method achieves comparable results to other popular classification methods such as Neural Network and Decision Tree which performs better in general.