南京大学学报（自然科学版）2013，Vol.49Issue(2)：150-158,9.
一种新的不平衡数据v-NSVDD多分类算法
A new unbalanced data v-NSVDD multiclass algorithm
摘要
Abstract
Based on analysis of the problems of state of the art support vector data description (SVDD) methods for multiclass classification,a new unbalance data v- NSVDD algorithm for multiclass classification is proposed. Inpired from the idea of v- SVM,SVDD with negative samples(NSVDD) and classification margin maximum principle, the method can not only effectively reduce the perturbances of noises and outliers, but also correct the problem existed in optimization problem in Ref[8],which considerably improves the generalization performance of the proposed algrithm. The training samples are weighted to deal with the unbalanced data classification problem, and also the weight coefficients can be conveniently calculated in theory according to the number of each class samples. Taking into account of complexity and non-linearity of the classfication samples in many practical applications, we extend the proposed linear algorithm to nonlinear cases by means of kernel trick. Since many nonlinear unbalanced data v - NSVDD models are needed for multiclass data and each model is seperately developed,the corresponding kernel parameters are very different. As a result,distences between a fixed sample and hypersphere centers in different reproduction kernel hilbert spaces (RKHSs) are not inconsistent with pratical distances. Furtherly,most existing multiclass classification methods are lack of rejection decision, which can impair the classification performance and reliability of decision. Relative distance measure is first proposed to achieve the consistent distances in RKHSs, and then a multiclass classification method is developed by combining relative distance with K-NN calssfication rule to deal with the above problems. The benchmark testing results show that the proposed method can provide lower classification errors and deal with unbalanced data problem.关键词
支持向量数据描述(SVDD)/样本类别不平衡/多分类/拒判/超球软边界Key words
support vector data description ( SVDD) / unbalanced data problem/ multiclass classification/ rejection decision/hypersphere soft boundary引用本文复制引用
刘小平,徐桂云,任世锦,杨茂云..一种新的不平衡数据v-NSVDD多分类算法[J].南京大学学报（自然科学版）,2013,49(2):150-158,9.基金项目
国家自然科学基金(60974056) （60974056）
