An Improved Support Vector Machines With Balanced Binary Decision Tree For Multi-class Classification

The support vector machines with balanced binary decision tree for multi-class classification has high recognition accuracy、fast identification characteristics, thus it becomes one of the most commonly used algorithms for solving multi-class classification problem. The main influence factors of the algorithm are the building scheme of support vector machines with balanced binary decision tree for multi-class classification, parameters of each decision face, feature weight vector of each sample’s eigenvector,this article focuses on these three aspects to improve the algorithm.It’s difficult to build an efficient balanced binary decision tree for support vector machine with balanced binary decision tree for multi-class classification. To solve this problem, this paper proposes a new scheme to build an efficient balanced binary decision tree which adopts a new scheme of distance measure between classes and introduces concepts of divisibility between classes and division factor. Finally, the contrast experiment shows that in case of invariance of training time complexity and recognition time complexity,the improved measure can efficiently enhance recognition accuracy.Currently,most of balanced binary decision tree for support vector machine with balanced binary decision tree for multi-class classification use Gussian kernel. The selection of parameter of Gaussian kernel can influence algorithm performance directly. Support vector machines with balanced binary decision tree for muli-class classification with different optimal Gaussian kernel’s parameter, so this algorithm need a fast、effective parameters selection algorithm for Gaussian kernel. Based on differentiability property of binary support vector machines objective function, this paper derivated a series of corollaries of derivated function of binary support vector machines objective function. From these deductions,this paper combine with deformation gradient descent method proposed a new Derivative Characteristics based parameters Selection Algorithm for Gaussian kernel named DC_SA, with optimum speed of convergence. Furthermore, this paper applies DC_SA to Gaussian kernel based support vector machines with balanced binary decision tree for multi-class classification. The contrast experiments of multi-class classification shows that support vector machines with DC_SA based balanced binary decision tree for multi-class classification can efficiently enhance recognition accuracy in case of invariance of training time complexity and recognition time complexity.Currently, most SVM algorithm handle feature dimension of eigenvector are the same, that is, to each feature dimension are using same weights value, but for a data set, each feature dimension in the eigenvector impact to SVM classification performance is different, some feature dimension are strong connected to the classification, some feature dimension are weak related to the classification and also some feature dimension not related to the classification. So if we use same of weights value for feature dimension, it will weaken SVM of classification performance, especially for support vector machines with balanced binary decision tree for multi-class classification. To overcome this problem, in this paper we propose the algorithm of feature dimension weight self-adapted, this algorithm takes weight to the feature dimension, outstanding the key feature dimension in the eigenvectors, effectively increases performance of the multi-class classification algorithm. The contrast experiments show that, introducing the RF_SA algorithm of support vector machines with balanced binary decision tree for multi-class classification’s recognition accuracy has been further improved in case of invariance of recognition time complexity.