A new method of rockburst prediction for imbalanced data categories using an improved XGBoost algorithm with enhanced interpretability

Rockburst prediction is critical for safety in deep underground engineering but is persistently hindered by complex geological conditions and severe class imbalance in monitoring data. To address these challenges, this study proposes a novel hybrid framework integrating Generative Adversarial Networks (GAN), the Guided Black-winged Kite Algorithm (GBKA), and Extreme Gradient Boosting (XGBoost). Specifically, GAN is employed to synthesize realistic minority-class samples to rebalance the dataset, while GBKA adaptively optimizes the hyperparameters of the XGBoost classifier. Based on 341 global rockburst cases with six geomechanical indicators, the proposed model achieves a classification accuracy of over 96% on the test set across four intensity levels (None, Light, Moderate, Strong), significantly outperforming conventional machine learning approaches. Furthermore, SHAP (SHapley Additive exPlanations) analysis is incorporated to provide model interpretability, revealing that maximum tangential stress (σθ) and the Stress Concentration Factor (SCF) are the dominant predictors, collectively contributing over 50% to the decision process. This finding aligns well with the stress-driven mechanisms of rockbursts. External validation on 12 independent mining cases in China demonstrates an 83.3% consistency rate with actual events. The proposed framework offers a highly accurate, interpretable, and robust solution for rockburst risk assessment, particularly in scenarios with limited and imbalanced data.