Partially replacing cement with mine tailings offers a sustainable strategy for solid waste resource utilization. As a cement admixture, the compressive strength of tailings-based cement materials serves as a critical performance indicator. Machine learning (ML) techniques offer high efficiency, cost-effectiveness, and superior predictive accuracy. However, variations in the chemical composition of tailings often introduce uncertainties into model predictions. Consequently, this study developed an integrated approach incorporating chemical composition and activation methods as input parameters. Four optimized ML models were deployed to predict the compressive strength of tailings-based cementitious materials. Multiple metrics were employed to evaluate model performance, which identified the PSO-XGBoost model as the superior predictive architecture. SHAP analysis revealed that mechanical grinding, NaOH concentration, and the proportions of gypsum and tailings were the primary features influencing compressive strength. Experimental validation yielded a low prediction error of 8.7%, confirming the model’s high predictive accuracy. This research establishes a robust framework for predicting the strength of tailings-based cementitious materials, providing a theoretical foundation for solid waste upcycling.