Abstract
This work presents a hybrid deep learning approach for identifying students who are likely to experience academic difficulties in virtual learning environments. The proposed framework is evaluated on the Open University Learning Analytics Dataset (OULAD) and combines two complementary types of information: temporal patterns of learner activity captured using Bidirectional Long Short-Term Memory (Bi-LSTM) networks and relatively stable student attributes modeled through a Multi-Layer Perceptron (MLP). To account for the fact that student engagement does not contribute equally across different phases of a course, a temporal attention mechanism is incorporated to emphasize the most informative periods over the 32-week learning timeline. The experimental results indicate that the hybrid model achieves stronger predictive performance than baseline approaches, with a ROC-AUC of 0.95 and a weighted F1-score of 0.90. Analysis of the attention weights suggests that both early engagement and activity toward the end of the course (Weeks 25-32) play an important role in predicting final outcomes. To address the imbalance between successful and at-risk learners, a cost-sensitive training strategy is adopted, resulting in a recall of 0.79 for the at-risk group. Overall, these findings suggest that integrating temporal behavioral signals with static student characteristics leads to more reliable risk prediction and provides a useful basis for data-informed academic support in online learning contexts.