[Objective] A landslide susceptibility evaluation framework that integrates the characteristics of Karst landforms was constructed， in order to reveal the dominant factors and their interaction mechanisms， and provide a theoretical basis for the precise control of landslide risks in Karst landform areas. [Methods] The study area was Guilin City， Guangxi Zhuang Autonomous Region. We innovatively integrated Karst characteristics to construct a geospatial database containing nine key factors. Five machine learning models-logistic regression （LR）， support vector machine （SVM）， random forest （RF）， gradient boosting machine （GBM）， and XGBoost-were used for landslide susceptibility modeling. Hyperparameters were optimized via a grid search， and the model performance was evaluated using accuracy and area under the receiver operating characteristic curve （AUC） metrics. The shapley additive explanations （SHAP） algorithm was applied to quantify the contributions of the various factors and their interactive effects. [Results] The ensemble models RF and XGBoost exhibited the best performance， achieving the highest accuracy （0.85 and 0.84， respectively） and AUC values （0.93 and 0.92， respectively）. All models showed a trend of a sharp increase in landslide density with decreasing land area. The RF model yielded the highest landslide density in the extremely high-risk zones （0.164 events/km²）. The SHAP analysis indicated that the groundwater content was the most influential Karst factor in most models. It also revealed the interactive effects of the normalized difference vegetation index， distance to rivers， and soil type. The ensemble models exhibited high consistency in terms of feature interpretation. [Conclusion] Integrating ensemble models such as RF with the SHAP framework significantly improves the accuracy and interpretability of landslide susceptibility mapping in Karst regions. The results of this study confirm the synergistic disaster-forming mechanism between the groundwater content and soil type.