Vector projection method for unclassifiable region of support vector machine

摘要

Unclassifiable region (UR) in conventional multi-classification support vector machine (MSVM) decreased the classification capacity and generalization ability of MSVM. To overcome the disadvantage, vector projection method (VPM) was presented. VPM first projects the samples in UR onto the line linking every two class centers, then computes the feature distance between each projecting point and corresponding class center. For one sample, the class with smaller feature distance will be voted one time and the sample belongs to the class which owns the most votes. Experimental results on synthetic and benchmark datasets show that VPM resolved the UR problem effectively and improved the classification capacity and generalization ability of MSVM.