一种LDA与SVM混合的多类分类方法  

Hybrid multi-class classification approach based on LDA and SVM

作　　者：张宏达[1] 王晓丹[1] 徐海龙[1] 

机构地区：[1]空军工程大学导弹学院,陕西三原713800

出　　处：《控制与决策》2009年第11期1723-1728,共6页Control and Decision

基　　金：陕西省自然科学基金项目(2007F19);空军工程大学导弹学院研究生学位论文创新基金项目(DY06102)

摘　　要：针对决策有向无环图支持向量机(DDAGSVM)需训练大量支持向量机(SVM)和误差积累的问题,提出一种线性判别分析(LDA)与SVM混合的多类分类算法.首先根据高维样本在低维空间中投影的特点,给出一种优化LDA分类阈值;然后以优化LDA对每个二类问题的分类误差作为类间线性可分度,对线性可分度较低的问题采用非线性SVM加以解决,并以分类误差作为对应二类问题的可分度;最后将可分度作为混合DDAG分类器的决策依据.实验表明,与DDAGSVM相比,所提出算法在确保泛化精度的条件下具有更高的训练和分类速度.To the problems that decision directed acylic graph support vector machines （DDAGSVMs） suffer from training a mass of SVMs and the error accumulation effect, a hybrid multi-class classification algorithm taking optimized linear discriminant analysis（LDA） and SVM as node classifiers is proposed. According to the characteristics of projecting high-dimension data to low-dimensional space, an optimized LDA classification threshold is derived. Then the linear separability of each pair of classes is defined with repect to the classification accuracy of the optimized LDA. The proposed algorithm trains SVMs for the dichotomies with relatively low linear separability, and then updates the separability matrix with classification accuracy of the SVMs. In the classification phase, the separability matrix is employed to decide the decision route of the DDAG. Finally, experiments show that, comparing with DDAGSVM, the proposed algorithm possesses higher training and classifying speed without the loss of generalization accuracy.

关 键 词：决策有向无环图 支持向量机 线性判别分析 分类阈值 可分性 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]