<p>Landslide susceptibility assessment in seismically active mountainous regions is essential for post-earthquake geohazard risk reduction, but traditional models are often limited by high-dimensional factor redundancy, model instability, and insufficient interpretability. This study takes Luding County, Sichuan Province, China, which was affected by the 2022 Ms 6.8 earthquake and is located in the Xianshuihe Fault Zone, as the study area. We propose an integrated framework that combines Principal Coordinate Analysis (PCoA) with a Bagged XGBoost Classifier. The framework is further supported by Tree-structured Parzen Estimator (TPE)-based hyperparameter optimization, Shapley Additive Explanations (SHAP) for model interpretation, and multi-dimensional uncertainty analysis. Results show that the PCoA-Bagged XGBoost Classifier achieved the best predictive performance among the tested models, with Precision, Recall, F1-score, Accuracy, and AUC values of 0.922, 0.943, 0.932, 0.932, and 0.984, respectively. Compared with XGBoost, Random Forest, and Bagged XGBoost, the proposed model increased AUC by 0.084, 0.094, and 0.102, respectively. The confusion matrix further confirmed its classification reliability, with 296 true negatives, 26 false positives, 16 false negatives, and 298 true positives. SHAP analysis indicates that distance from fault and distance from river are the dominant controls on landslide susceptibility, contributing 28.7% and 24.3%, respectively. Uncertainty analysis shows that the model remains stable when landslide sample retention exceeds 60% (AUC ≥ 0.974), whereas performance declines when retention falls below 40%. The proposed framework provides a robust, interpretable, and uncertainty-aware tool for landslide susceptibility assessment in earthquake-affected mountainous regions.</p>