Bidirectional sampling method for imbalanced data
Junjie Shi, Deyu Song, Shengyao Zheng, Yueming Hu, Shuangshuang Chen, Fengque Pei · 2023
Traditional over-sampling and under-sampling algorithms suffer from overfitting and high noise when unbalanced data classes are in the sample set. To improve the performance of the data classifier, this study proposes a SMOTECU algorithm combining SMOTE and ClusterCentroids under-sampling. It absorbs the advantages of both algorithms and avoids generating or rejecting excessive samples in the dataset, effectively reducing the harmful effects of overfitting and noise. We experiment with 16 unbalanced standard datasets combining three classifiers: RF, RBFNN, and RBFSVM. By comparing three evaluation metrics: F1-score, AUC, and running time, the results demonstrate that the performance of the SMOTECU-based random forest model is better, and compared with SMOTE and ClusterCentroids, SMOTECU can effectively avoid overfitting and save running time.