Greedy-AutoML: A novel greedy-based stacking ensemble learning framework for assessing soil liquefaction potential

Abstract

Automated machine learning (AutoML) is a generic term for a specific approach to machine learning (ML) area that tries to automate the end-to-end process of employing repetitive ML tasks for real-world problems. In recent years, the AutoML framework, which is the subject of an increasing number of research articles, has become a potential approach for developing complicated ML models without human experience and support. Although existing techniques on AutoML have yielded promising results, research in this field is immature, and new approaches should be developed gradually. This study describes a novel AutoML framework to predict soil liquefaction potential problem based on stacking ensemble learning (SEL) combined with a greedy search algorithm. A special AutoML framework, called Greedy-AutoML, is presented that automatically produces an optimized ML model for predicting on a supervised classification task. The general concept of the proposed AutoML framework consists of three main steps: data preparation, greedy feature selection, and greedy stacking ensemble. Furthermore, the Greedy-AutoML framework has been published on a user-friendly web-based platform for testing or trial purposes. To highlight the capability of this AutoML application, Greedy-AutoML is applied to predict the liquefaction potential of soils using three well-known datasets (i.e., CPT — cone penetration test, SPT — standard penetration test, and Vs — shear wave velocity test) collected from previously published research. The results are assessed based on different performance matrices, namely Accuracy (Acc), Kappa, Precision, Recall, and F1-Score. Experiments with datasets from existing case histories with varying distribution and features showed that the proposed greedy based SEL method achieved an Acc of 98% for the CPT and SPT datasets, while the achieved Acc for the Vs dataset was about 99%.