Abstract:
In today's educational landscape, identifying at-risk students and providing personalized, adaptive
learning experiences remain critical challenges. Traditional Learning Management Systems (LMS)
lack dynamic tools for predicting student performance, adapting learning pathways, and
recommending content based on individual needs. This project addresses these gaps by developing
an LMS that utilizes advanced AI techniques to enhance grade prediction, personalize content
recommendations, and support early interventions for at-risk students.
A hybrid approach combining a Transformer-based LSTM autoencoder and Explainable AI (XAI)
techniques was implemented. The autoencoder leverages multi-head self-attention and positional
encoding to capture temporal learning behaviour patterns, and it employs anomaly detection via
reconstruction error to identify at-risk students. The inclusion of XAI further enhances
transparency by elucidating the model’s prediction rationale, thereby aiding educators in
understanding which features contributed to a student’s risk status. Comprehensive data
preprocessing, feature selection, and hyperparameter tuning supported by cross-validation and
ablation studies optimized the model’s accuracy and adaptability.
In parallel, a dual-path neural network merges student and content metadata through element-wise
feature interactions to generate real-time, adaptive recommendations. SHAP-based XAI highlights
key drivers behind suggestions, fostering trust. This module supports proactive guidance by
tailoring resources to individual needs, enhancing the system’s capacity for personalized academic
support.
Early testing showed promising results, with an accuracy rate of 85% and an AUC-ROC of 0.92.
The LSTM-based grade prediction model effectively identified at-risk students, achieving a
balanced F1 score of 0.85. These results suggest the system's potential to enhance student success
through timely, personalized learning interventions.
