Enhancing soil creep index prediction accuracy using optimized XGBoost and LightGBM models with bio-inspired optimization algorithms: SHAP- and LIME-based explanation methods

Soil creep, as a form of long-term deformation, significantly influences the performance of foundations and earth structures. As this deformation develops over time, it can lead to major geotechnical issues such as slope failures, differential ground movement, and excessive settlement. Consequently, accurate estimation of the soil creep index is essential for informed geotechnical decision-making. To address this, the primary aim of this study is to develop precise models for estimating the soil creep index using modern machine learning methods. This study develops fourteen predictive models, including seven based on XGBoost and seven based on LightGBM. These include the standard XGBoost and LightGBM models, as well as their optimized variants enhanced using grey wolf optimizer (GWO), whale optimization algorithm (WOA), bat algorithm (BAT), dragonfly algorithm (DOA), crayfish optimization algorithm (COA), and grasshopper optimization algorithm (GOA). To assess the reliability of creep index prediction, the models were evaluated under four separate sets of input variables. In addition, SHAP, LIME, and SelectKBest methods were applied to explore the influence of individual input factors on the prediction outcomes. Each model’s effectiveness was assessed through four principal statistical indicators: the mean absolute error (MAE), root mean square error (RMSE), correlation coefficient (R), and Nash-Sutcliffe efficiency (NSE). Based on the RMSE values, the top three hybrid XGBoost models were XGBoost-GOA (RMSE = 0.002259), XGBoost-GWO (RMSE = 0.002273), and XGBoost-COA (RMSE = 0.002316). Among the hybrid LightGBM models, the best performers were LightGBM-BAT (RMSE = 0.003813), LightGBM-GWO (RMSE = 0.003884), and LightGBM-COA (RMSE = 0.003885). For the standalone models, XGBoost (RMSE = 0.003084) outperformed CART (RMSE = 0.003398) and LightGBM (RMSE = 0.003972). The testing results indicate that XGBoost-based models provide a reliable and accurate approach for predicting the soil creep index.