Abstract:

Unclassifiable region (UR) is one of the primary disadvantages in conventional multi-classification support vector machine (MSVM). To overcome the shortage and enhance the classification capacity and generalization ability of MSVM, a sample density method (SDM) is presented. SDM first constructs a hypersphere which considers the sample falling into the UR as a center and a certain threshold as radius. Then, sample density for each class in the hypersphere is computed and the class with the largest sample density is labelled  for the sample. Experimental results on synthetic datasets and benchmark datasets show that SDM  eliminates the UR in conventional MSVM and improves the classification performance of MSVM effectively.