Unifying Data Refinement and Fusion Strategies: A Cutting-Edge Methodology for E-Learning Performance Optimization

The surge in the e-learning market has compelled the need for intelligent predictive models to minimize the gap between performance analysis and personalized learning systems to improve student performance. With the wide growth of the e-learning market, there is a great demand for intelligent predictive models to bridge the gap between performance analysis and personalized learning systems to boost student performance. Yet, traditional prediction methods are not effective due to issues like class imbalance, outliers, irrelevant features, and limited model interpretability. This research introduces an advanced framework to predict student academic performance based on an educational data mining dataset, Students' Academic Performance Dataset (xAPI-Edu-Data), comprising 480 instances, 16 attributes, and multi-variate integer and categorical features of the e-learning environment and educational data mining. The proposed framework involves class imbalance handling using Synthetic Minority Oversampling Technique (SMOTE), outlier removal using Interquartile Range (IQR), feature scaling and data standardization using Z-score normalization. A novel hybrid feature selection method TRIFEX is proposed to select the most influencing features to the student performance by combining ANOVA F-statistics, Recursive Feature Elimination (RFE) and Lasso regularization. The Logistic Regression, Decision Tree, and K-Nearest Neighbor (KNN) classifiers are used in the study. The hyperparameter optimization is done using Randomized Search CV, Grid Search CV, and Optuna to increase the efficiency and generalization power of the model. In addition, a voting, based ensemble model is created to fuse the virtues of individual classifiers for good prediction. Experimental results have shown that the proposed ensemble model is more accurate with 98.99% accuracy, 99.00% F1-score and 0.1005 RMSE as compared to the conventional predictive models. The results suggest that the suggested method has substantial potential to increase the accuracy of prediction, explainability of features and individualized learning support in contemporary e-learning environments.