不均衡数据混合取样分类算法  

A classification algorithm based on mixed sampling for imbalanced dataset

作　　者：杜红乐[1] 张燕[1] 

机构地区：[1]商洛学院数学与计算机应用学院,陕西商洛726000

出　　处：《燕山大学学报》2015年第2期158-164,共7页Journal of Yanshan University

基　　金：陕西省自然科学基金资助项目(2014JM2-6122);陕西省教育厅科技计划资助项目(12JK0748);商洛学院科学与技术研究项目(13sky024)

摘　　要：针对不均衡数据分类决策面偏移导致少数类识别率较低的问题,提出一种混合取样算法.首先计算类样本数的比值K;然后分别在多数类和少数类中随机选取一个样本,计算该样本的K-1近邻,以K个样本的中心作为新样本;再对剩余的样本重复上面操作,直到所有样本都被处理;最后所得新样本与原少数类样本共同构成新的训练集.该算法在改变样本密度的同时保持了原样本的空间分布,实验结果表明该算法能够提高SVM在不均衡数据下的分类性能,尤其是少数类的分类性能.In order to solve the problem of the lower accuracy of minority class caused by classification hyper plane shifting,a mixed sampling algorithm for imbalanced data classification is proposed.First,the ratio of the numbers of majority class and minority class, K is calculated.Then,a sample is chosen randomly from majority class and minority class,and K-1 nearest neighbors of the sample are calculated,the center of above K samples is taken as a new sample.Above processing is repeated until all samples are processed. The new generated samples and original minority class are put together as a new training dataset.At the same time, the sample den？sity is changed and the sample distribution in the feature space is kept.Experiment results show this proposed algorithm can improve the classification performance of SVM for imbalanced dataset,especially for the minority class.

关 键 词：支持向量机 过取样 不均衡数据集 欠取样 K 近邻 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]