Imbalanced rock burst assessment using variational autoencoder-enhanced gradient boosting algorithms and explainability

We conducted a study to evaluate the potential and robustness of gradient boosting algorithms in rock burst assessment, established a variational autoencoder (VAE) to address the imbalance rock burst dataset, and proposed a multilevel explainable artificial intelligence (XAI) tailored for tree-based ensemble learning. We collected 537 data from real-world rock burst records and selected four critical features contributing to rock burst occurrences. Initially, we employed data visualization to gain insight into the data's structure and performed correlation analysis to explore the data distribution and feature relationships. Then, we set up a VAE model to generate samples for the minority class due to the imbalanced class distribution. In conjunction with the VAE, we compared and evaluated six state-of-the-art ensemble models, including gradient boosting algorithms and the classical logistic regression model, for rock burst prediction. The results indicated that gradient boosting algorithms outperformed the classical single models, and the VAE-classifier outperformed the original classifier, with the VAE-NGBoost model yielding the most favorable results. Compared to other resampling methods combined with NGBoost for imbalanced datasets, such as synthetic minority oversampling technique (SMOTE), SMOTE-edited nearest neighbours (SMOTE-ENN), and SMOTE-tomek links (SMOTE-Tomek), the VAE-NGBoost model yielded the best performance. Finally, we developed a multilevel XAI model using feature sensitivity analysis, Tree Shapley Additive exPlanations (Tree SHAP), and Anchor to provide an in-depth exploration of the decision-making mechanics of VAE-NGBoost, further enhancing the accountability of tree-based ensemble models in predicting rock burst occurrences.