H-SVMs的构造方法  

Construction methods for H-SVMs

作　　者：闫志刚[1,2] 杜培军[1] 

机构地区：[1]中国矿业大学环境与测绘学院,徐州221116 [2]中国矿业大学深部岩土力学与地下工程国家重点实验室,徐州221008

出　　处：《东南大学学报（自然科学版）》2009年第S1期204-209,共6页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(40802061;40401038);国家重点基础研究发展计划(973计划)资助项目(2007CB209406);中国博士后科学基金资助项目(20080441081);江苏省博士后科研资助计划资助项目(0901057C)

摘　　要：通过推导H-SVMs推广能力的模型,得出H-SVMs的推广能力与样本类别数、空间分布、容量、树结构等有关,且保证高优先级结点的推广性能是提高H-SVMs性能的有效途径。根据分析结果,提出了一种基于SVM最大间隔分类、最小间隔聚类构造H-SVMs的新方法。利用SVM的分类间隔作为分类、聚类指标,通过Top-down和Bottom-up两种途径混合构造H-SVMs,其中,最大间隔分类采用Top-down策略,在各结点依次选择最大间隔的SVM,将输入样本按类别分为2类;最小间隔聚类采用Bottom-up策略,在各结点依次选择最小间隔的SVM,将输入样本按类别两两聚类。从UCI数据库中选取多类数据进行测试,实验结果验证了该方法的有效性,说明所构造的H-SVMs具有较好的、稳定的推广性能。By deducing the formula for the generalization power of hierarchical support vector machines(H-SVMs),it can be concluded that the generalization power of H-SVMs is related to the count,the distribution,the capacity of the input samples and H-SVMs' structure.And the guarantee of the generalization ability of the SVM node,whose position is located at a high level,can improve the performance of H-SVMs.Then,a novel method for building H-SVMs is put forward.The separation distances of SVMs are regarded as the indices for classifying and clustering.Through the top-down and bottom-up routes,the input samples are classified by maximal separation distances and clustered by minimal separation distance.The approach of classification by maximal separation distance can select the SVM whose separation margin is maximal based on the top-down route,and dichotomize the input samples according to their categories at each node.The approach of clustering by minimal separation distance can select the SVM whose separation margin is minimal based on the bottom-up route,and hierarchically cluster every two input samples according to their categories at each node.Several multiclass data sets from UCI database are selected to test this method.The experimental results show that the method is practicable,and the corresponding H-SVMs structure has a preferable and stable generalization ability.

关 键 词：H-SVMs 分类树 最小间隔聚类 最大间隔分类 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]