Improved prediction of soil shear strength using machine learning algorithms: interpretability analysis using SHapley Additive exPlanations

The soil’s shear strength is an important parameter that is used frequently throughout the design phase of construction. The conventional method of calculating shear strength in a laboratory is more expensive and time-consuming. This study presents an attempt to develop models for predicting soil shear strength with improved accuracy, particularly Extreme Gradient Boosting (XGBoost), Gradient Boosting (GB), Adaptive Boosting (AdaBoost), and Categorical Boosting (CatBoost). The Coefficient of determination (R 2 ), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Deviation (MAD) indices were used to validate each of the developed models. The analysis of the results demonstrates that the AdaBoost model achieved a better prediction performance with R 2 = 0.99794 and lowest values of RMSE = 0.00400, MAE = 0.00080, MAPE = 0.24390 and MAD = 0.00080 followed by the CatBoost model with R 2 = 0.99651, RMSE = 0.00521, MAE = 0.00429. MAPE = 1.33450 and MAD = 0.00429 in the training phase when compared to previous models such as multivariate adaptive regression splines and support vector regression published in the literature. In addition, SHapley Additive Explanations analysis elucidates that the liquidity index has the greatest influence on soil shear strength, followed by wet density.