A Bidirectional GRU-Based Framework for Early Prediction of Student Academic Failure in E-Learning Environments

Early prediction of students’ academic performance in e-learning systems plays a crucial role in reducing dropout rates, improving the learning process, and designing targeted educational interventions. Despite recent advances in machine learning and deep learning methods, many existing studies are either unable to effectively model the temporal dependencies in students’ behavioral data or require a large portion of semester-long data to make reliable predictions. In this paper, a framework based on a Bidirectional Gated Recurrent Unit (Bi-GRU) network is proposed for the early prediction of at-risk students. The proposed method utilizes behavioral data from the Open University dataset and relies solely on interaction records collected during the first 42 days of the course. It extracts and analyzes temporal activity patterns of students. During the preprocessing stage, multiple statistical and temporal features are extracted, including cumulative activity counts, moving averages, activity change rate, number of active days, and time since last activity. The data are then structured into sequential formats, normalized, and enhanced before being fed into a Bi-GRU-based architecture, which effectively learns temporal dependencies and behavioral dynamics of students. To address data imbalance, a weighted loss function is employed, and early stopping is used to prevent overfitting. Experimental results demonstrate that the proposed method achieves an accuracy of 67.37%, an AUC of 73.54%, and an F1-score of 63.99%, indicating its effectiveness in identifying at-risk students at early stages of the course. The findings suggest that modeling student behavioral sequences using deep temporal learning approaches can provide an effective tool for early warning systems in educational environments.