一种新的基于聚类的试探性SMOTE算法

New tentative SMOTE algorithm based on clustering

针对传统过采样算法中常常出现的生成噪声点、数据分布边缘化、未增强足够特征的问题,提出了一种新算法:试探性少数类过采样技术(tentative synthetic minority over-sampling technique,TSMOTE)。该算法先将少数类样本进行K-means聚类,然后创建正类安全水平等指标,运用试探性的思想,放出试探点求出每个少数类样本对应的警戒点,获取最适合新样本生成的空间区域,最后在簇心和警戒点之间进行合成少数类过采样技术(synthetic minority over-sampling technique,SMOTE),确保新样本的生成质量。在12个公开数据集上的大量实验表明:TSMOTE算法可以有效提高分类器对少数类样本和整体数据集的分类性能。

Aiming at the phenomenon of generating noise points,marginal data distribution,and missing minority features that often occur in traditional oversampling algorithms,a new algorithm named by the Tentative Synthetic Minority Over-sampling Technique(TSMOTE)is proposed.The algorithm first performs K-means clustering of minority samples,and then introduces indicators such as positive safe level.Relying on tentative ideas,it releases tentative points to find the warning line corresponding to each minority sample,and obtains the most suitable new sample generation,Synthetic Minority Over-sampling Technique(SMOTE)is performed between the cluster center and the warning line to ensure the generation quality of new samples.A large number of experiments on 12 public data sets show that the TSMOTE algorithm can effectively improve the classification performance of the classifier on minority samples and overall data sets.