Method based on support vector machine and sequential backward selection for seismic liquefaction potential evaluation

In the paper, the support vector machine (SVM) is utilized to evaluate the earthquake-induced site liquefaction potential, and an optimization algorithm based on cross validation and sequential backward selection(SBS) is proposed to improve the generalization ability of the classifier for seismic liquefaction potential evaluation(SLPE). Usually, the accuracy of SLPE using the SVM varies greatly when the training dataset and test dataset change, so the classifier is not reliable enough in practice. Because cross validation is more convincing for evaluating the classifier performance in machine learning, the algorithm in the paper tries to reduce the maximum error of cross validation through adopting SBS to determine the input variables of the SVM. The performance of the classifier is assessed by the area under the curve (AUC) on the basis of confusion matrix. As shown by data validation, the algorithm can reduce the maximum error of cross validation and the variation of accuracy in SLPE while maintaining good performance of the classifier. In conclusion, a method that can improve the reliability of SVMs for classification in SLPE is put forward in the paper.