基于L1范数凸包数据描述的多观测样本分类算法.pdfVIP

第34 卷第1 期 电 子 与 信 息 学 报 Vol.34No.1 2012 年1 月 Journal of Electronics Information Technology Jan. 2012 基于L1 范数凸包数据描述的多观测样本分类算法 胡正平* 王玲丽 (燕山大学信息科学与工程学院 秦皇岛 066004) 摘 要：为建立高维空间样本分布的最佳覆盖为目标来实现覆盖分类，该文提出基于L1 范数凸包数据描述的多观 测样本分类算法。首先对训练集的每个类别以及测试集的多观测样本分别构造凸包模型，这样多观测样本的分类就 转化为凸包模型的相似性度量问题。若测试集的凸包模型与训练集无重叠，采用L1 范数距离测度进行凸包模型之 间的相似性度量；若有重叠，采用L1 范数距离测度进行收缩凸包(reduced convex hulls)之间的相似性度量。然后 采用最近邻准则作为多观测样本的分类决策。在3 个数据库上进行的实验结果，表明该文提出方法对于多观测样本 分类具有可行性和有效性。 关键词：模式识别；凸包；L1 范数距离测度；最近邻分类；多观测样本 中图分类号：TP391.41 文献标识码： A 文章编号：1009-5896(2012)01-0194-06 DOI: 10.3724/SP.J.1146.2011.00545 The Classification Algorithm of Multiple Observation Samples Based on L1 Norm Convex Hull Data Description Hu Zheng-ping Wang Ling-li (School of Information Science and Engineering, Yanshan University, Qinhuangdao 066004, China) Abstract: In order to construct a high-dimensional data approximate model in the purpose of the best coverage of the distribution of high-dimensional samples, the classification algorithm of multiple observation samples based on L1 norm convex hull data description is proposed. The convex hull for each class in the train set and multiple observation samples in the test set is constructed as the first step. So the classification of multiple observation samples is transformed to the similarity of convex hulls. If the test convex hull and every train hull are not overlapping, L1 norm distance measure is used to solve the similarity of convex hulls. Otherwise, L1 norm distance measure is used to solve the similarity of reduced convex hulls. Then the nearest neighbor classifier is used to solve the classification of multiple o