Interpretable prediction for seismic failure modes of RC columns via key parameter selection and cluster‐improved hybrid sampling

Abstract

Accurate prediction of seismic failure modes in reinforced concrete columns is critical to performance‐based seismic design and vulnerability assessment of building and bridge structures. However, conventional machine learning (ML) models are hindered by subjective parameter selection and severe class imbalance. This study proposes an interpretable framework integrating two novel methods: interpretable key parameter (IKP) selection, which couples Recursive Feature Elimination, Explainable Boosting Classifier, and mechanical principles to objectively identify influential parameters; and cluster‐improved hybrid sampling (CHS), which combines clustering, over‐sampling, and under‐sampling to generate high‐quality balanced datasets. Validated against 543 experimental specimens, the proposed IKP‐CHS framework improves average precision, F1‐score, Kappa, and Matthews correlation coefficient by approximately 5%–6% over the original data and 8%–12% over conventional sampling methods. Notably, recall for the minority failure modes (flexure‐shear and shear) increases by 33% and 10%, respectively, compared to traditional ML models. A user‐friendly graphical interface further enables practical deployment. The framework provides a robust, interpretable, and data‐driven solution for seismic failure mode identification, effectively addressing class imbalance and parameter subjectivity.