RQSP-SMOTE: a multi-linear interpolation oversampling method based on regular quadrilateral scoring mechanism with perturbation

Abstract Imbalanced data classification is a common task across various fields, and oversampling is a key strategy in this context. As an effective oversampling method, SMOTE has gained widespread recognition. It generates new samples by leveraging existing data samples through specific construction strategies. However, the basic SMOTE is not suitable for complex data feature spaces. Therefore, this paper proposes a novel multi-linear interpolation oversampling method based on regular quadrilateral scoring mechanism with perturbation (RQSP-SMOTE). The RQSP-SMOTE algorithm exploits the geometric properties of regular quadrilaterals and introduces perturbations to establish a new scoring mechanism. It dynamically selects samples and performs multi-linear interpolations in the original dimensional space to synthesize new samples. Meanwhile, it avoids the dependency on the k-nearest neighbor method. Comparative experiments with other improved SMOTE algorithms, integrated with multiple classifiers and evaluation metrics, show that the RQSP-SMOTE method achieves overall superior performance. These results indicate that RQSP-SMOTE effectively enhances classification performance on imbalanced datasets, yielding superior outcomes after oversampling.