Abstract
Background
Predicting student performance in online learning is difficult due to class imbalance and limited model interpretability. At-risk students are fewer than high performers, biasing predictions, and methods like SMOTE fail to preserve temporal patterns. Although black-box models are accurate, they lack transparency for actionable insights.
Objectives
This study proposes a deep learning framework combining LSTM networks and attention mechanisms to address these issues using the OULAD dataset. LSTMs capture temporal dependencies, while attention improves interpretability by emphasising key features. Advanced resampling mitigates class imbalance for robust at-risk student detection.
Methods/Approach
The methodology applies the KDD framework to process data, uncover patterns, and build models that predict student success risk, ensuring efficient data handling, robust modelling, and actionable insights to improve retention.
Results
The BiLSTM-RNN achieved the best performance, effectively capturing temporal dependencies and attaining the highest accuracy, precision, recall, and F1-score.
Conclusions
The findings support more effective and targeted interventions in online education, offering valuable insights for research and practice.