邻域自适应SMOTE算法研究  

Research on neighborhood adaptive SMOTE algorithm

作　　者：王芳 吴文通 张立立 马瑞 徐文星[1,2] Wang Fang;Wu Wentong;Zhang Lili;Ma Rui;Xu Wenxing(College of Information Engineering,Beijing Institute of Petrochemical Technology,Beijing 102617,China;Academy of Artificial Intelligence,Beijing Institute of Petrochemical Technology,Beijing 102617,China)

机构地区：[1]北京石油化工学院信息工程学院,北京102617 [2]北京石油化工学院人工智能研究院,北京102617

出　　处：《计算机应用研究》2021年第6期1673-1677,共5页Application Research of Computers

基　　金：北京市属高校青年拔尖人才培育计划资助项目(CIT&TCD201704048);北京市教委—市自然基金资助项目(KZ202110017025)。

摘　　要：针对SMOTE(synthetic minority over-sampling technique)等基于近邻值的传统过采样算法在处理类不平衡数据时近邻参数不能根据少数类样本的分布及时调整的问题,提出邻域自适应SMOTE算法AdaN_SMOTE。为使合成数据保留少数类的原始分布,跟踪精度下降点确定每个少数类数据的近邻值,并根据噪声、小析取项或复杂的形状及时调整近邻值的大小;合成数据保留了少数类的原始分布,算法分类性能更佳。在KEEL数据集上进行实验对比验证,结果表明AdaN_SMOTE分类性能优于其他基于近邻值的过采样方法,且在有噪声的数据集中更有效。To solve the problem of traditional oversampling algorithms based on neighbor values such as SMOTE(synthetic minority over-sampling technique)that the nearest neighbor parameters cannot be adjusted in time according to the distribution of minority samples when dealing with imbalanced data,this paper proposed a neighborhood adaptive SMOTE algorithm AdaN_SMOTE.In order to keep the original distribution of the minority class in the synthetic data,this algorithm determined the neighbor value of each minority class data by tracking the precision decline point,and adjusted the size of the neighbor value in time according to noise,small disjunctions or complex shapes.Thus,the synthetic data retained the original distribution of the minority classes,which made the algorithm classification performance better.Experimental comparison and verification on the KEEL datasets show that the classification performance of AdaN_SMOTE is better than other oversampling methods based on nearest neighbor values,especially in noisy datasets.

关 键 词：类不平衡 数据分布 自适应邻域大小 精度下降点 人工合成少数类过采样 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]