A temporal attention-based hybrid deep learning model for student performance and academic risk prediction

Area of Science:

• Educational Technology
• Artificial Intelligence in Education
• Learning Analytics

Background:

• Virtual learning environments present challenges in identifying students at risk of academic difficulty.
• Traditional methods may not fully capture the dynamic nature of student engagement over time.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning framework for predicting academic risk in online learners.
• To integrate temporal learner activity with static student attributes for enhanced prediction accuracy.

Main Methods:

• A hybrid deep learning model combining Bidirectional Long Short-Term Memory (Bi-LSTM) for temporal data and Multi-Layer Perceptron (MLP) for static attributes.
• Incorporation of a temporal attention mechanism to weigh the importance of different course phases.
• Utilized the Open University Learning Analytics Dataset (OULAD) for evaluation.
• Employed a cost-sensitive training strategy to address class imbalance.

Main Results:

• The hybrid model achieved a ROC-AUC of 0.95 and a weighted F1-score of 0.90, outperforming baseline approaches.
• Attention analysis highlighted the significance of early and late-course engagement (Weeks 25-32) for predicting outcomes.
• A recall of 0.79 for the at-risk group was achieved using the cost-sensitive strategy.

Conclusions:

• Integrating temporal behavioral signals with static student characteristics improves the reliability of academic risk prediction.
• The findings support data-informed academic interventions in online learning contexts.
• The model provides a valuable tool for proactive student support in virtual education.