基于聚类的少数类样本采样方法

针对处理不平衡数据集的分类问题,SMOTE通过在相邻样例间线性插值实现少数类样例过采样。但SMOTE插值的结果是样例密集的地方依然相对密集,样例稀疏的地方依然相对稀疏,影响分类性能。针对该问题本文提出一种基于聚类的过采样方法-C-SMOTE。该方法首先将少数类样例聚成多个簇,再以簇为单位结合SMOTE方法产生新样例。实验结果表明,C-SMOTE既保证了数据集整体分类准确率,又能提高少数类分类精度。

Clustering Based Oversampling Approach for Minority Class Data

Abstract SMOTE is a popular oversampling method in dealing with the classification of imbalanced data sets, which produces instances by linear interpolation between existing minority instances. The main disadvantage is that after SMOTE, it＇s still intensive where itg intensive and still sparse where it~ sparse in sample space. So we propose a cluster - based oversampling method （ C - SMOTE） to overcome this disadvantage. In this method, the minority class samples will be clustered into several groups firstly, and then groups are used as units to generate new samples with SMOTE. Experimental results show that our method works well in improving the classification ac- curacy for minority class without reducing the overall accuracy.