EMERGING TRANSFORMER-POWERED PERSONALIZED LEARNING METHODOLOGY

Emerging Knowledge Tracing (KT) models, particularly deep learning and attention-based Knowledge Tracing, have shown great potential to enable personalized learning. One of the most important tasks in educational technology is knowledge tracing (KT), which models a student’s current state of knowledge based on previous interactions in order to predict how well they will perform in the future. Existing methods mainly focus on immediate past interactions and fail to fully capture the hierarchical dependencies between individual knowledge concepts and do not account for how new knowledge builds upon a sequence of prior knowledge concepts, known as knowledge concept routes, that can be critical to advance the understanding of student learning outcomes. To address this, in our work, we propose an innovative attention-based KT method by effectively incorporating the educational domain knowledge of knowledge concept routes for the given curriculum. By constructing a Learning Relevance Matrix that identifies relationships between questions based on their shared concept routes, this approach masks attention scores for completely unrelated questions, ensuring that only relevant question pairs contribute to the model's predictions. This method is implemented based on the Attentive Knowledge Tracing model architecture and leveraging the XES3G5M and DBE-KT22 datasets, with rich auxiliary information for knowledge concept routes and dependencies, to evaluate and compare the performance of our proposed method to the seven State-of-the-art (SOTA) KT models. Empirical results demonstrate that our proposed model significantly improves predictive performance, achieving a 0.97 AUC score compared to 0.86 for the baseline model without the concept route-based attention mechanism on the XES3G5M dataset, and a 0.98 AUC score as compared to 0.94 on the DBE-KT22 dataset. Furthermore, our method shows a 7% improvement in accuracy on the XES3G5M dataset and a 5% improvement on the DBE-KT22 dataset. These improvements indicate that incorporating long-term dependencies between knowledge concepts significantly improves prediction accuracy and can lead to better interpretability of student performance, contributing to the development of personalized learning systems. This work contributes to the field of educational data mining by addressing key limitations in existing KT models and providing a more cognitively aligned approach to modeling student learning.