基于混合重采样的非平衡数据SVM训练方法  

SVM training with imbalanced dataset based on mixed resampling

作　　者：郭亚伟[1] 白治江[1] 

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《微型机与应用》2016年第12期52-54,58,共4页Microcomputer & Its Applications

摘　　要：针对传统的SVM算法在非平衡数据分类中分类效果不理想的问题,提出一种基于分类超平面和SMOTE过采样方法(HB_SMOTE)。该方法首先对原始训练样本集使用WSVM算法找到分类超平面,然后按一定标准剔除负类中被错分的样本、靠近分类超平面的样本以及远离分类超平面的样本。在UCI数据集上的实验结果表明:与RU_SMOTE等重采样方法相比,HB_SMOTE方法对正类样本和负类样本都具有较高的分类准确率。The classification result of classical SVM algorithm in the case of unbalanced data set is unsatisfactory. Therefore,a class hyperplane based SMOTE methods（ HB_SMOTE） is presented. The new method firstly finds the class hyperplane by using WSVM on the original imbalanced dataset,then according to a specific criterion,the negative class is discarded into the misclassified samples,the samples close to hyperplane and the samples far away to the hyperplane. Finally the experiment results on the UCI dataset show the new method performs in higher accuracy,compared to the RU_SMOTE and other similar algorithms.

关 键 词：非平衡数据集 SMOTE 分类超平面 SVM 混合重采样 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]