基于数据关系的SVM多分类学习算法

提出一种基于数据关系(Data Relationship,DR)的多分类支持向量机(Support Vector Machine,SVM)学习算法(Multi-Classification SVM Algorithm Based on Data Relationship,DR-SVM).DR-SVM算法根据每类数据的关系(如向量积等)获取子学习嚣的冗余信息,从而优化多分类器组,然后通过经典的SVM算法训练分类器组.算法在简化分类器组的同时可对多类数据分类问题获得满意的泛化能力,在标准数据集上的实验结果表明,与经典的SVM多分类方法相比,DR-SVM具有更好的泛化性能,尤其对单个类别精度要求较高的数据尤其有效.

A Multi-Classification SVM Algorithm Based on Data Relationship

A multi-class support vector machine(SVM)algorithm was introduced based on data relationship(DR-SVM).Through extracting redundant information of subclassifiers based on relationship between different class data(such as inner product of vectors and so on),the DR-SVM model can reduce the number of subclassifiers.Then the multi-classifiers can be trained by traditional SVM.In so doing,the obtained model can be simplified and the satisfactory generalization performance can be reached at same time.The experiment results on benchmark datasets demonstrate that comparing with several traditional multi-class SVM approaches,the DR-SVM possesses better performance.Especially,it is more efficient for some data processing problems like the predicting precision of individual class should be not under a threshold.