Abstract
The problem of imbalanced data classification is a prominent and challenging research topic in the field of data mining and machine learning. Numerous studies have demonstrated that synthetic minority oversampling technique (SMOTE) and its variants are widely adopted methods for addressing imbalanced data training. However, the performance of SMOTE and its variants can be affected by noise. Additionally, most existing techniques used to handle noise in SMOTE variants involve directly deleting noisy samples, which may lead to class re-imbalance and deviation of the decision boundary. Furthermore, SMOTE and its variants do not guarantee the diversity of synthetic samples. Motivated by these limitations, this study aims to propose a novel oversampling method named TRPS-DER to tackle class-imbalanced classification problems. TRPS-DER utilizes triangular region pre-sampling for synthesizing minority class samples and employs differential evolution resampling for filtering out noise. The primary advantage of TRPS-DER include that (a) it generates minority class samples by interpolation of triangular region, thereby augmenting diversity of synthesize samples; and (b) it employs differential evolution for resampling generated samples, effectively filtering out noise and improving classification performance. Extensive experimental results demonstrate that TRPS-DER significantly outperforms other competitive SMOTE-based oversampling methods across 24 imbalanced datasets in terms of Gmean, BACC, AUC.