Deep Knowledge Tracing Using Bi-LSTM and Attention Mechanisms for Enhanced Mastery Prediction in Personalized E-Learning

This study proposes a Deep Knowledge Tracing (DKT) model that integrates Bidirectional Long Short-Term Memory (Bi-LSTM) networks with an attention mechanism to improve mastery prediction and interpretability in personalized e-learning environments. Using a dataset of 1,250 learners and 482,903 LMS interaction logs spanning 38 unique skills, the model was trained to predict next-response correctness and estimate latent knowledge states across dynamic learning sequences. The average sequence length of 74.2 interactions per learner and correctness mean of 0.63 indicate substantial variability in learning trajectories, emphasizing the need for robust temporal modeling. Results show that the proposed model achieves superior predictive performance compared with baseline approaches. Specifically, it reaches an accuracy of 0.781, AUC-ROC of 0.854, F1-score of 0.767, log-loss of 0.451, and ECE of 0.031, outperforming standard DKT (AUC-ROC 0.823) and LSTM models (AUC-ROC 0.806). The proposed Bi-LSTM + Attention approach provides both high predictive accuracy and explainable outputs, enabling more informed instructional interventions, personalized recommendations, and early-risk detection in large-scale e-learning environments. These findings affirm the methodological and pedagogical value of combining bidirectional sequence modeling with attention-driven interpretability for next-generation adaptive learning systems.