Effect of Feature Selection Technique on the Pile Capacity Predicted Using Machine Learning

Pile capacity is an important issue in geotechnical engineering, with substantial practical implications. The practice currently relies on a few traditional mechanics-based design methods; however, with the recent advancement in machine learning (ML), many studies started investigating its diverse applications such as predicting pile capacity. The accuracy of the predicted capacity is associated with the quality of the information used as input features. This emphasizes the significance of the feature selection process. In this study, the effect of seven feature selection techniques on the performance of nine machine learning models are investigated using a dataset of 481 piles. The outcome was then compared to the performance of traditional design methods. It was concluded that using the support vector regression ML model combined with sequential feature selection technique resulted in a better performance in terms of precision and the mean absolute percentage error over available methods.