This paper evaluated 10 different machine learning algorithms for predicting wall displacement of nailing systems. The input parameters were all dimensionless and real-case data, including H/H0, z/H, c/γH, φ/φo, L/H, ShSv/At, qs/γH, and Ka. The output was the normalized horizontal displacement. The models considered include linear approaches (Linear, Lasso, Ridge Regression), distance-based methods (K-Nearest Neighbors, Support Vector Regression), and ensemble models (Random Forest, AdaBoost, XGBoost, CatBoost, Gradient Boosting). Using an 80/20 train/test split and feature standardization, each of the above models was validated using 5-fold cross-validation and systematic hyperparameter optimization. A feature importance analysis showed that the normalized depth of measurement and nail spacing density are the major predictors accounting for about 48% of the predictive power. Among the models studied, Gradient Boosting proved to be the most efficient, with a cross-validated R2 score of 0.725 ± 0.098, an RMSE of 0.680 ± 0.122, and an MAE of 0.445 ± 0.079 after tuning. The results on the test dataset (R2 = 0.648, RMSE = 0.765) demonstrated generalization ability without any evident signs of overfitting. The tuned Gradient Boosting model was a good example of an efficient, interpretable, and robust alternative to traditional approaches, offering immediate practical advantages in geotechnical engineering.
