Abstract: The raw material sources of alkali-activated materials (AAMs) are more complex than those of Portland cement, and their reaction mechanisms are different. Therefore, the influencing factors for performance design and prediction should be set based on the principle of alkali activation. This study explored the feasibility of using alkali activation reactivity of raw materials as machine learning input features to predict the compressive strength of slag-based AAMs, comparing them with traditional features such as oxide composition and pozzolanic activity. First, the pozzolanic and alkali reactivity of granulated ground blast furnace slag (GGBS), fly ash(FA), metakaolin(MK), and silica fume(S) were characterized using the strength method and the "XRF-alkali leaching-XRF" method. Then, 102 paste samples were prepared by mixing the four raw materials and sodium silicate activator in different proportions, and their 28 d compressive strengths were measured as output features. Based on the material reactivity and proportions, two sets of input features with six labels each were constructed. Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Stacking models were employed to predict compressive strength, with model performance evaluated using R², RMSE, and MAE. Furthermore, SHAP values were applied for interpretable analysis of the models. The results indicate that the alkali activation reactivity can refine the traditional oxide composition features into reactive oxides and insoluble components with opposite effects. These parameters also replace the ambiguous influence of pozzolanic activity on the compressive strength of AAMs. Consequently, using alkali activation reactivity as input features significantly improves the predictive accuracy of all models. Among them, XGBoost demonstrated superior learning and generalization capabilities, with the R² on the test set increasing from 0.84 to 0.87.