基于局部相关性的kNN分类算法  

A kNN Classification Algorithm Based on Local Correlation

作　　者：邓振云[1] 龚永红[1,2] 孙可[1] 张继连[1] 

机构地区：[1]广西师范大学广西多源信息挖掘与安全重点实验室,广西桂林541004 [2]桂林航天工业学院,广西桂林541004

出　　处：《广西师范大学学报（自然科学版）》2016年第1期52-58,共7页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金资助项目(61573270;61263035;61363009);国家973计划项目(2013CB329404);广西自然科学基金资助项目(2012GXNSFGA060004;2015GXNSFCB139011);中国博士后科学基金资助项目(2015M570837);广西多源信息挖掘与安全重点实验室开放基金资助项目(MIMS13-08)

摘　　要：kNN算法作为一种简单、有效的分类算法,在文本分类中得到广泛的应用。但是在k值(通常是固定的)的选取问题上通常是人为设定。为此,本文引入了重构和局部保持投影(locality preserving projections,LPP)技术用于最近邻分类,使得k值的选取是由样本间的相关性和拓扑结构决定。该算法利用l1-范数稀疏编码方法使每个测试样本都由它的k(不固定)个最近邻样本来重构,同时通过LPP保持重构前后样本间的局部结构不变,不仅解决了k值的选取问题,并且避免了固定k值对分类的影响。实验结果表明,该方法的分类性能优于经典kNN算法。As a simple and effective classification algorithm,kNN algorithm is widely used in text classification.However,the k value（usually fixed）is usually set by users.For this purpose,the reconstruction and locality preserving projections（LPP）technology is introduced into the nearest neighbor classification,which makes the selection of the kvalue to be determined by the correlation between the samples and the topology structure.The algorithm uses l1-norm sparse coding method to reconstruct the test sample by its k（not fixed）nearest neighbor samples and LPP keeps the local structure of the sample after the reconstruction,which not only solves the problem of choosing kvalue,but also avoids the influence of fixed k value on classification.Experimental results show that the classification performance of the proposed method is better than that of the classical kNN algorithm.

关 键 词：KNN 保局投影 重构 稀疏编码 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]