基于L1范数凸包数据描述的多观测样本分类算法

The Classification Algorithm of Multiple Observation Samples Based on L1 Norm Convex Hull Data Description

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。

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 observation samples.Experiments on three types of databases show that the proposed method is valid and efficient.