Stacking model based on six base classifiers to improve prediction of soil liquefaction: a multi-dataset study

Although the existing machine learning methods have achieved satisfactory prediction accuracy on specific datasets, they failed to perform well on other datasets. To overcome the limitation, a novel ensemble algorithm called the stacking model is proposed to improve earthquake-induced liquefaction potential of soil. The stacking model was constructed by combining the predicted results of six base classifiers using the weight voting method and a secondary classifier (logistic regression). The six base classifiers are random forest, logistic regression, support vector machine, naïve bayes, back propagation neural network and K-nearest neighbour, respectively. The grid search algorithm was used to search the hyperparameters, and the area under the curve (AUC) value was utilised as the basis to obtain the optimal hyperparameters. The predicted result of the stacking model has been compared with that of six-base classifier and a previous study. The result demonstrated that the proposed model outperforms the base classifier in all three datasets in terms of a variety of evaluation metrics. Also, the proposed model improved the highest accuracy achieved by previous studies on three datasets by 5.2%, 3% and 4.7%, respectively. Further, this study investigated the importance of input parameters so as to interpret the complicated relationship between liquefaction potential and input parameters.