Accurately predicting student learning outcomes is a critical step toward enhancing educational strategies and supporting personalized learning. This study investigates the effectiveness of six machine learning (ML) classification algorithms—Random Forest, Logistic Regression, Decision Tree, K-Nearest Neighbors, AdaBoost, and XGBoost—for forecasting student performance. To address class imbalance within the datasets, the Synthetic Minority Oversampling Technique (SMOTE) was applied, ensuring robust model training. The experimental framework evaluated the classifiers using two separate datasets, containing 525 and 493 records respectively, and employed multiple cross-validation methods to ensure generalizability. Among the tested models, XGBoost consistently delivered superior performance. On the first dataset, XGBoost achieved an accuracy of 96.83%, sensitivity of 98.13%, specificity of 94.12%, precision of 97.22%, F1 score of 97.67%, and an AUC of 99.10%. For the second dataset, it yielded 95.95% accuracy, 98.0% sensitivity, 91.67% specificity, 96.08% precision, 97.03% F1 score, and 98.21% AUC. To enhance model interpretability, SHAP (SHapley Additive exPlanations) was integrated into the framework, offering valuable insights into the influence of input features on prediction outcomes. The proposed approach not only demonstrates high predictive accuracy but also emphasizes explainability, offering a practical and cost-effective tool for stakeholders in education and potentially transferable to other domains requiring early-stage predictive modeling.