Machine learning-based Debris Flow Susceptibility Mapping (DFSM) has emerged as an effective approach for assessing debris flow likelihood, yet its application faces three critical challenges: insufficient reliability of training samples caused by biased negative sampling, opaque decision-making mechanisms in models, and subjective susceptibility mapping methods that lack quantitative evaluation criteria. This study focuses on the Yalong River basin. By integrating high-resolution remote sensing interpretation and field surveys, we established a refined sample database that includes 1,736 debris flow gullies. To address spatial bias in traditional random negative sampling, we developed a semi-supervised optimization strategy based on iterative confidence screening. Comparative experiments with four tree-based models (XGBoost, CatBoost, LGBM, and Random Forest) reveal that the optimized sampling strategy improved overall model performance by 8%-12%, with XGBoost achieving the highest accuracy (AUC = 0.882) and RF performing the lowest (AUC = 0.820). SHAP-based global-local interpretability analysis (applicable to all tree models) identifies elevation and short-duration rainfall as dominant controlling factors. Furthermore, among the tested tree-based models, XGBoost optimized with semi-supervised sampling demonstrates the highest reliability in debris flow susceptibility mapping (DFSM), achieving a comprehensive accuracy of 83.64% due to its optimal generalization-stability equilibrium.