Abstract: To enable rapid prediction of the thermal performance of cement-based grouting composites for shallow geothermal applications and to identify the key influencing factors, 330 valid records were collected from multiple literature sources. A thermal-conductivity database containing 12 input variables related to mixture composition, pore structure, moisture condition, and bulk density was established. Gaussian process regression (GPR), support vector regression (SVR), random forest (RF), CatBoost, and XGBoost were compared, and the responses of the XGBoost model were interpreted using SHAP dependence analysis, PDP–ICE, and two-dimensional partial dependence analysis. The results show that XGBoost achieved a test-set coefficient of determination (R2) of 0.9282 and an RMSE of 0.2780 W·m?1·K?1, which was close to the RMSE of GPR. Considering predictive accuracy, nonlinear representation capability, and interpretability, XGBoost was selected as the principal model. The SHAP results indicate that total porosity makes the largest contribution to the model output, followed by the water-to-binder ratio, sand-to-binder ratio, granular-phase-to-binder ratio, and bulk density. The PDP–ICE results show that the predicted thermal conductivity decreases markedly when total porosity increases to approximately 34%–36%; the water-to-binder ratio and degree of saturation exhibit stage-dependent responses, whereas the sand-to-binder ratio and conductive filler-to-binder ratio generally show positive responses. The two-dimensional partial dependence results further reveal joint responses for bulk density–degree of saturation, water-to-binder ratio–sand-to-binder ratio, and total porosity–sand-to-binder ratio. These findings provide a reference for thermal-conductivity prediction, key-factor identification, and mixture screening of cement-based grouting composites for shallow geothermal applications.