Abstract: The Loess Plateau is a region highly susceptible to landslide occurrence. Accurately identifying landslide-prone areas and understanding the key driving factors are critical for disaster prevention, mitigation, and large-scale infrastructure planning in the region. This study focuses on the Loess Plateau and selects 13 evaluation factors based on multicollinearity analysis. Three optimized models were developed: the Extreme Gradient Boosting (XGBoost) model, the Frequency Ratio-Coupled XGBoost (FR-XGBoost) model, and the Frequency Ratio-Coupled Random Forest (FR-RF) model. From the perspective of model interpretability, this study further investigates the driving mechanisms and interactions among key factors. The results indicate that the coupled models (FR-XGBoost and FR-RF) outperform the single XGBoost model in predictive accuracy, with the FR-XGBoost model achieving the highest AUC value of 0.968. Annual average rainfall, soil erodibility, and slope were identified as the primary driving factors for landslides in the study area, contributing 26.59%, 20.80%, and 14.66% to the model output, respectively. Additionally, partial dependence plots (PDPs) revealed nonlinear interaction effects among these key variables, showing that the combination of slope with rainfall or soil properties significantly enhances landslide susceptibility within specific ranges.