Predictive Modeling of Student Performance in Adaptive Learning Using Ensemble Machine Learning and Behavioral Analytics

This study develops a comprehensive predictive modeling framework to analyze student performance within adaptive learning environments using ensemble machine-learning techniques. A multi-source dataset of interaction logs, behavioral metrics, and assessment histories was processed into 22 engineered features, including engagement ratio, performance stability, hint-usage patterns, and time-on-task trends. Three ensemble models Random Forest, Gradient Boosting Machine (GBM), and XGBoost were trained and evaluated through 5-fold cross-validation. Results show that XGBoost achieved the strongest predictive performance with an accuracy of 0.91, precision of 0.90, recall of 0.89, F1-score of 0.89, and an AUC-ROC of 0.94, outperforming GBM (accuracy 0.89) and Random Forest (accuracy 0.87). Cross-validation stability analysis indicates minimal metric variance (accuracy range: 0.90–0.92), confirming strong generalization across heterogeneous learners. Error analysis revealed that Medium-performing learners produced the widest residual range (up to ±0.25), reflecting unstable engagement patterns and mixed performance behaviors. SHAP explainability further identified engagement ratio (0.34), average quiz score (0.28), and time-on-task (0.18) as the strongest global predictors, while hint-usage ratio and performance-stability features contributed smaller but meaningful influences.  Overall, the findings highlight the strength of ensemble machine-learning models, particularly XGBoost, in capturing nonlinear learning patterns and supporting evidence-driven adaptive instruction. The integration of explainability and temporal modeling provides valuable pedagogical insights, offering a transparent and scalable foundation for next-generation intelligent tutoring and personalized learning systems.