一种基于改进BTS的多类非平衡分类的集成学习方法  

An ensemble classification method for multi-class imbalanced datasets

作　　者：汤志亚[1,2] 赵亮[1,2] 杨玲[1,2] 甄小琼[1,2] 杨志鹏[1,2] 

机构地区：[1]成都信息工程学院,四川成都610225 [2]中国气象局大气探测重点开放实验室,四川成都610225

出　　处：《商丘师范学院学报》2015年第6期30-34,共5页Journal of Shangqiu Normal University

基　　金：科技部公益性行业(气象)科研专项课题(GYHY201106047)支持

摘　　要：提出一种适用于多类不平衡数据的集成学习方法,以解决多类样本分布不均衡问题.首先,利用合成少类样本的过采样技术(Synthetic Minority Over-sampling Technique,SMOTE)得到一组类别平衡的训练集.然后,对每个训练集采用二叉树支持向量机(SVM of Binary Tree,BTS)进行训练,最后,采用Bagging进行集成.通过5组UCI测试数据表明该算法在Gmean参数上比SMOTEBagging算法提高2.55%.To solve the problem of multi- class imbalanced datasets classification,an ensemble classification method for multi- class imbalanced datasets is presented. Firstly,we get balanced datasets by using the SMOTE algorithm. Then,we utilize the binary tree of SVMs to get the models for every training dataset. Finally,we make an ensemble with bagging method for the models. The result showed that compared with SMOTEBagging,our method improved 2. 55% on the Gmean value to five UCI multi- class imbalanced benchmark datasets.

关 键 词：多类不平衡分类 集成方法 二叉树支持向量机 SMOTE算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]