Support vector machines for multi-class pattern recognition based on improved voting strategy

The improved voting strategy for pairwise classification of multi-class support vector machine (MSVM) is proposed. The new voting strategy can increase recognition accuracy and resolve the unclassifiable region problems caused by conventional pairwise classification. The improved voting value equals to the traditional voting value plus the tuning function. For the data in the classifiable regions, the classification results using improved voting strategy are the same as that using the traditional one. However, the data in the unclassifiable region can be determined by the tuning function. By computer simulations using four UCI data sets, the superiorities of the presented multi-class strategy are demonstrated.