Accurate prediction of site amplification effects is crucial for the seismic design of structures, especially in soft soil sites where seismic amplification is more pronounced. This study compares machine learning models (XGBoost, Random Forest, and Deep Neural Networks) with traditional empirical models (SS14) in predicting site amplification effects across different seismic periods. A dataset from the KiK-net strong-motion observation network was utilized, integrating soil profile features, seismic motion characteristics, and topographic information. To address the inherent class imbalance in the dataset, a hybrid data augmentation strategy was employed, combining down-sampling for weak-motion samples and augmentation of moderate- and strong-motion samples, thereby enhancing the robustness of the training model. Additionally, feature engineering was conducted, introducing features such as shear stiffness contributions, shear wave velocity gradients, and weighted averages to better characterize site conditions. The results show that XGBoost outperforms both Random Forest and Deep Neural Networks across all periods, providing superior predictive accuracy and generalization ability. In contrast, although Deep Neural Networks perform well with the training dataset, they exhibit underfitting in capturing complex site amplification effects, especially at higher frequencies. The empirical model SS14 shows higher prediction errors at shorter periods, primarily due to its inability to account for nonlinear soil behavior. Feature attribution analysis using SHAP values further validates the importance of Vs-weighted and Vse parameters, while highlighting the limited contribution of spectral acceleration (Sa) alone. This paper demonstrates the effectiveness of machine learning models, particularly XGBoost, in predicting site amplification effects and emphasizes the importance of feature selection, model architecture, and data augmentation strategies in seismic hazard assessment.