摘要
由于SMOTE算法插值时没有对边界和噪声样本做差异性处理,提出将邻域分布应用在SMOTE算法上的改进算法E_SMOTE。其核心是将正类样本按K近邻信息分为安全集和非安全集。对安全集按照SMOTE插值;对非安全集,在插值前探察其K近邻候选点的M近邻分布(简称M扩展近邻),从而控制新样本的合成区域,提升平衡数据集的抗噪性。在6个UCI数据集上训练SVM分类器,与SMOTE和SMOTE_NCL算法相比,E_SMOTE算法取得更高的F_value和G_mean值,表明分类器的总体分类性能有明显提高。
When interpolating new samples, the Synthetic Minority Over-sampling Technique（SMOTE） algorithm does not distinguish the noisy and boundary samples from normal ones. Presents an improved algorithm E-SMOTE that applies neighborhood distribution on SMOTE algorithm. The key of the new algorithm is to split the original positive samples into security and non-security according to K-nearest neighbor information. Then applies the SMOTE interpolation to the security set. Meanwhile, for the non-security set, further explores the distribution of M-neighbors（referred to as M-extended neighbors） for each of the K-nearest neighbor to decide whether a new sample will be interpolated. Consequently, the influence of the noisy samples is overcome through controlling the synthetic area of the new sample. The experiments for Support Vector Machine（SVM） classifier on six UCI datasets shows that, compared with the SMOTE and SMOTE_NCL oversampling algorithms, the E_SMOTE oversampling algorithm achieves higher value of F_value and G_mean, which indicates that the overall classification performance of the new classifier is significantly improved.
作者
宋艳
白治江
SONG Yan;BAI Zhi-jiang(College of Information Engineering,Shanghai Maritime University,Shanghai 201306)
出处
《现代计算机》
2018年第10期34-38,共5页
Modern Computer
