Interpretable Gradient Boosting Ensemble Model for Predicting Seismic Failure Modes of RC Columns Based on Class-Balanced Datasets

To address the unsatisfactory generalization ability and inherent interpretability of traditional models for class-imbalanced datasets, a novel interpretable gradient boosting ensemble (IGBE) model was proposed to intelligently predict seismic failure modes of reinforced concrete columns using the improved hybrid sampling method and the explainable boosting classifier. A hybrid sampling method was proposed to generate the class-balanced datasets by combining the clustering, the over-sampling and the under-sampling algorithms. The generalization ability of the proposed model was validated compared to traditional models and 543 sets of experimental data. Global and local explanations for the seismic failure modes prediction were further investigated. Analysis results show that the IGBE model exhibits satisfactory generalization performances in both training and test sets. For the test set, the IGBE model can improve the accuracy and Kappa coefficient by about 5–9% and 10–16%, respectively, compared with traditional sampling methods. Meanwhile, the misclassified rate of the IGBE model is reduced by about 16–49% compared with typical empirical models. The precision and recall of the IGBE model for the flexure-shear failure are increased by about 28% and 12% on average, while the precision and recall of the shear failure are increased by about 10% and 18% on average compared with traditional machine learning models.