XDSVM: A Machine Learning Classification Model for SMART Learning Analytics

Abstract

This study proposes a hybrid XDSVM framework for multi-class student performance prediction by integrating a Support Vector Machine (SVM) and a compact Deep Neural Network (DNN). Boruta feature selection identifies 20 key predictors from engagement and interaction data, capturing over 70% normalized feature gain. The DNN uses a 32→16 architecture with batch normalization and dropout between 0.2-0.4, while the SVM employs an RBF kernel optimized using Ray Tune for hyperparameter tuning. An adaptive weighted soft-voting and neural fusion meta-network combine outputs based on confidence and class imbalance. Evaluated on 29,850 instances across 8 performance classes, the model achieves 97.31% test accuracy, with precision 0.975, recall 0.9791, and F1-score 0.9771. The framework improves robustness, interpretability, and generalization, while providing risk scoring and personalized learning insights. Overall, XDSVM demonstrates strong predictive capability and scalability for educational analytics systems with potential real-world deployment in intelligent tutoring environments and adaptive learning platforms successfully validated.