Integrating Bayesian and Particle Swarm Optimization in Ensemble Machine Learning for Enhanced Prediction of Soil Liquidity Index and Shear Strength
摘要
This study presents a comprehensive exploration of the development of diverse stacking machine learning models for predicting the Soil Liquid Index (LI) and Soil Undrained Shear Strength (Su). The study utilizes a wide range of datasets, including over 1550 data points for LI and 950 for Su, collected from various research studies. Individual models such as Multilayer Perceptron (MLP), Random Forest (RF), and AdaBoost regressors were evaluated alongside ensemble models, including Simple Averaging (SA), Weighted Averaging (WA), and stacking-SVR. The optimization process, employing Particle Swarm Optimization (PSO) and Bayesian Optimization (BO), was used to fine-tune hyperparameters. The findings suggest that the AdaBoost regressor outperforms other individual models in LI prediction, while RF is the most proficient model for predicting Su. Ensemble techniques, particularly WA, significantly enhanced model performance, achieving remarkable coefficient of determination (R2) values in Su (0.9913 for BO, 0.9961 for PSO) and LI prediction (0.9624 for BO, 0.9338 for PSO) during testing. Combining base learners and stacking with Support Vector Regression (SVR) further improved performance, surpassing other models. In the optimization comparison, BO demonstrated superior LI prediction, while PSO excelled in Su prediction, highlighting the distinctive strengths of each technique. SHAP analysis further identified the depth of cone penetration (d) as the most influential factor for both LI and Su, thereby enhancing model interpretability and demonstrating strong alignment with established geotechnical principles.