SIRSYN: Improved Synthetic Sampling for Imbalanced Data Using SIR

Imbalanced data, where class labels in a training dataset are significantly skewed, often reduces the prediction accuracy for minority classes when using traditional algorithms. To address this, various oversampling techniques like SMOTE have been proposed, but they often generate minority data in majority class regions. We propose an improved method, SIRSYN, which applies the Sampling Importance Resampling (SIR) method to existing synthetic data generation techniques. This reduces the risk of generating data in inappropriate locations. Using 60 imbalanced datasets from the KEEL repository, we compared SIRSYN with 13 existing methods. SIRSYN achieved superior performance, G-means and F1 scores, indicating its effectiveness in enhancing oversampling techniques for imbalanced classification tasks.