Enhancing Student Engagement Prediction Using a Hybrid Gradient Boosting Approach

Abstract

The most critical issue in educational technology is the ability to predict how students will engage in the e-learning setting, but the implications of this prediction on the student learning outcomes, rate of course completion, and effectiveness of the institution has become one of the most critical issues in modern teaching technology. The paper introduces a new hybrid gradient boosting model, which will be a combination of the complementary strengths of XGBoost and LightGBM into a single framework of ensemble models in predicting the level of student engagement in online learning platforms. Based on the data of behavioral interaction extracted out of the EdNet-KT4 LMS Student Engagement Dataset large-scale hierarchical educational dataset which is the most complete and richest of the four EdNet hierarchical levels, containing 297,915 student files, with a compressed size of 1.2GB and an uncompressed size of 6.4GB. The proposed model is systematically compared to five separate unique baseline algorithms, including Random Forest, XGBoost, LightGBM, Artificial Neural Network, and Logistic Regression. In addition we compare our work with other state of art on the same dataset. The proposed hybrid achieves 96.00% accuracy, precision of 1.000, F1-Score of 0.960, and AUC of 0.982. The findings show that synergistic performance improvement with the strategic combination of the second-order gradient optimization of XGBoost and the leaf-wise growth and the histogram-based efficiency of LightGBM is synergistic as neither algorithm can achieve the synergistic improvement performance by itself. The implications of the results in practical use in implementation of real-time and scalable engagement monitoring systems in online educational institutions have significant practical value.