少样本条件下基于K⁃最近邻及多分类器协同的样本扩增分类  

Sample augmentation and classification method based on KNN and multi⁃classifier collaboration with limited samples

作　　者：陈伟杰 郑成勇[1] 蔡圣杰 罗智玉 CHEN Weijie;ZHENG Chengyong;CAI Shengjie;LUO Zhiyu(School of Mathematics and Computer Science,Wuyi University,Jiangmen 529000,China)

机构地区：[1]五邑大学数学与计算科学学院,广东江门529000

出　　处：《现代电子技术》2022年第15期123-127,共5页Modern Electronics Technique

基　　金：广东省自然科学基金资助项目(2018A030313063)。

摘　　要：针对少样本条件下的分类问题,提出一种基于K⁃最近邻及多分类器协同的训练样本扩增分类框架。首先利用少量标记样本对多个分类器进行初步训练,并在整个样本空间中搜索出每个标记样本的K个最近邻;然后利用初步训练好的分类器,对每个标记样本的K个最近邻进行分类,若某个最近邻被多数或全体分类器判为具有与其标记样本相同的类,则将该最近邻判别为与其标记样本同类,并将其添加至该标记样本所属类的扩展训练样本集,利用扩展训练样本集再次对各分类器进行训练;最后利用再次训练过的多个分类器对剩余未标记样本进行基于投票的分类判决。在多个基准测试数据库上的对比实验结果表明,在少标记样本条件下,所提算法能显著提升分类器的分类精度。A training sample augmentation and classification framework with K⁃nearest neighbor(KNN)and multi⁃classifier collaboration is proposed to eliminate the difficulties in the classification of the limited samples.A small quantity of labeled samples are used for preliminary training of multi⁃classifier,and the K nearest neighbors of each labeled sample are searched in the entire test sample space.And then,the preliminarily⁃trained classifiers are used to classify the K nearest neighbors of each labeled sample.If a certain nearest neighbor is judged to have the same class as the labeled sample by the majority or all of the classifiers,then the nearest neighbor is judged to share the same class as its labeled sample and is added to the extended training sample set of the class to which the labeled sample belongs.After this,each classifier is trained again with the extended training sample set.Finally,the retrained multiple classifiers are used to make a voting⁃based classification decision on the remaining unlabeled samples.The contrastive experimental results on the multiple benchmark databases show that the proposed algorithm can significantly improve the classification accuracy under the condition of limited labeled samples.

关 键 词：样本扩增分类 K⁃最近邻 多分类器协同 少样本 投票法 半监督分类 样本筛选 

分 类 号：TN911.1-34[电子电信—通信与信息系统]