Abstract: To address the issues of anomalous sample characteristics and imbalanced sample sizes across different rock burst intensity levels, a combined method integrating the Stochastic Outlier Selection (SOS) algorithm and the Borderline−SMOTE algorithm for minority class oversampling is proposed. This approach aims to enhance dataset quality and improve the accuracy of prediction models. A rock burst intensity prediction model was developed based on a Support Vector Machine (SVM) optimized by the Tianji Horse Racing Optimization (THRO) algorithm. First, 315 collected rock burst datasets were processed using the SOS algorithm to remove outliers across all intensity levels. Subsequently, the Borderline−SMOTE algorithm was applied to generate synthetic samples for minority classes near the decision boundary. The datasets before and after processing were used to train and validate four machine learning models: AdaBoost, XGBoost, Random Forest, and SVM. Compared with the original dataset, the data processed by the combined SOS and Borderline−SMOTE method improved the prediction accuracy of the four models by 16.74%、8.51%、6.78% and 14.88%, respectively. The THRO−SVM model was further compared with models including WOA−SVM, WOA−RF, SSA−SVM, and IDBO−RF. The results indicate that the THRO−SVM model achieves the highest accuracy and stability. Engineering case studies, such as that in the Shaiqi River Phosphate Mine area, validate the reliability of the proposed rock burst intensity prediction model combining SOS, Borderline−SMOTE, and THRO−SVM algorithms.