An Explainable Hybrid Deep Learning Framework for Engagement-Aware Academic Performance Prediction in Adaptive Learning Systems

Adaptive learning systems are increasingly recognized as essential tools for enhancing student engagement and improving academic outcomes in data-driven educational environments. Traditional one-size-fits-all instructional approaches often fail to account for individual behavioral patterns, the level of motivation and the learning preferences, which results in a decrease in engagement and an under-performing achievement. This paper suggests a data-driven multi-task deep learning model that is aimed at studying student engagement and predicting academic performance at the same time. The behavioral, psychological, demographic and academic features of 14,003 anonymized student records were used. The engagement levels were calculated through the K-Means clustering by key behavioral indicators without specifying labels allowing objective segmentation of the learners. The development of a shared-representation neural network architecture was subsequently created to predict the categories of engagement and final grades. As shown in the experimental results, the predictive performance is strong with an engagement classification accuracy of about 99 % and the final grade prediction of about 98 % with weighted precision, recall and F1-scores of over 0.97. The values of ROC-AUC of above 0.97 are a confirmation of high discriminative ability. The suggested framework enables early detection of the at-risk learners and enables customized instructional interventions under intelligent adaptive learning systems.