Geometric Construction Method of Linear SVM Multi-class Classifier
Ying Tang, Yingzhen Li · Jisuanji gongcheng · 2012
(Abstract )A new method to construct multi-class classifier based on linear SVM is proposed in the paper. Its major procedures include: to form interval space polygon among point sets by subtracting operation of convex hulls, to extract polygon axes and then extend to construct the classification boundaries. The method can avoid problems like blind area in decision-making and imbalance data sets like traditional multi-class classification ways of One-Against-All(OAA) and One-Agianst-One(OAO) encounter. Simulation test results show that classification boundaries constructed by the method can realize the minimum risk and the maximum interval space among point sets, thus can be seen as an embodiment of the optimal classification lines of multi-class point sets. (Key words ) ) ) )Support Vector Machine(SVM); optimal classification line; convex hull; Delaunay triangulation; polygon axis; multi-class classification