Data Importance Measurement Based on Sampling Region Information for Oversampling

Sampling is commonly employed to tackle the classification of imbalanced data, with the Synthetic Minority Oversampling Technique (SMOTE) being the most widely used sampling method. In recent years, multiple variations of SMOTE, including the clustering & SMOTE based oversampling methods are proposed. Nevertheless, these methods often neglect the calculation of importance measurement for minority class samples while generating noise, boundary, or overlapping samples in some cases. To solve such problems, we propose a novel importance measurement based on sampling region information (SRI) in minority class samples by using clustering. The measurement focuses more on important samples. Based on the measurement, we also propose an oversampling method called KRISMOTE. The method not only reduces noise, but also clarifies the classification boundaries, effectively enhancing the classification performance of imbalanced data. Experimental results over the publicly available KEEL dataset demonstrate that the proposed KRISMOTE method outperforms other popular oversampling algorithms.