OALDPC: oversampling approach based on local density peaks clustering for imbalanced classification

SMOTE has been favored by researchers in improving imbalanced classification. Nevertheless, imbalances within minority classes and noise generation are two main challenges in SMOTE. Recently, clustering-based oversampling methods are developed to improve SMOTE by eliminating imbalances within minority classes and/or overcoming noise generation. Yet, they still suffer from the following challenges: a) some create more synthetic minority samples in large-size or high-density regions; b) most fail to remove noise from the training set; c) most heavily rely on more than one parameter; d) most can not handle non-spherical data; e) almost all adopted clustering methods are not very suitable for class-imbalanced data. To overcome the above issues of existing clustering-based oversampling methods, this paper proposes a novel oversampling approach based on local density peaks clustering (OALDPC). First, a novel local density peaks clustering (LDPC) is proposed to partition the class-imbalanced training set into separated sub-clusters with different sizes and densities. Second, a novel LDPC-based noise filter is proposed to identify and remove suspicious noise from the class-imbalanced training set. Third, a novel sampling weight is proposed and calculated by weighing the sample number and density of each minority class sub-cluster. Four, a novel interpolation method based on the sampling weight and LDPC is proposed to create more synthetic minority class samples in sparser minority class regions. Intensive experiments have proven that OALDPC outperforms 8 state-of-the-art oversampling techniques in improving F-measure and G-mean of Random Forest, Neural Network and XGBoost on synthetic data and extensive real benchmark data sets from industrial applications.