A Cluster-based Undersampling Technique for Multiclass Skewed Datasets

Abstract

Imbalanced data classification is a demanding issue in data mining and machine learning. Models that learn with imbalanced input generate feeble performance in the minority class. Resampling methods can handle this issue and balance the skewed dataset. Cluster-based Undersampling (CUS) and Near-Miss (NM) techniques are widely used in imbalanced learning. However, these methods suffer from some serious flaws. CUS averts the impact of the distance factor on instances over the majority class. Near-miss method discards the inter-class data within the majority of class elements. To overcome these flaws, this study has come up with an undersampling technique called Adaptive K-means Clustering Undersampling (AKCUS). The proposed technique blends the distance factor and clustering over the majority class. The performance of the proposed method was analyzed with the aid of an experimental study. Three multiminority datasets with different imbalance ratios were selected and the models were created using K-Nearest Neighbor (kNN), Decision Tree (DT), and Random Forest (RF) classifiers. The experimental results show that AKCUS can attain better efficacy than the benchmark methods over multiminority datasets with high imbalance ratios.