Machine learning-based seismic capability evaluation for school buildings

The dataset of the school building seismic capability evaluations from the National Center of Research on Earthquake Engineering (NCREE) in Taiwan opens up the possibility of developing a rapid screening model using machine learning techniques. In this study, an imbalanced dataset composed of 8951 records collected over the past two decades using detailed pushover analysis for seismic capability evaluation are used. Five resampling techniques and three classifiers are developed and applied to address classification performance. We found that the combination of Cluster-based Synthetic Minority Oversampling (CBS) with the Random Forest (RF) classifier achieves the best performance (0.861) using the weighted average of the F1 score as the performance metric. However, the combination of Cluster-based Oversampling (COS) with the RF classifier achieves the best performance (0.853) for the area under the receiver operating characteristic curve (ROC-AUC) performance metric. In addition, three of the top five features are basic structural information from the design drawings. This implies that the seismic capability is heavily related to the initial design.