Attempt to reduce the effect of biased data-set on ground-motion prediction using machine learning

Abstract

Previous study [Kubo 2018] tried to construct a predictor of ground-motion index using a random forest method and strong-motion data recorded in Japan. However, the data-set is very biased and there are few strong ground-motion records. This causes the underestimation of the predictor for strong ground-motions. To overcome this problem, in this study, we suggest two approaches: one is the weighting of train data, and the other is the hybrid method integrating the conventional ground motion prediction equation and a machine learning approach. The verification using test data indicates that the hybrid method can largely improve the underestimation, although the underestimation still remains in predicting very strong groundmotions (>1000 gal).