Моделі відстеження знань в адаптивних навчальних системах: порівняння BKT і SAKT

Abstract

This study compares Self-Attentive Knowledge Tracing (SAKT) and Bayesian Knowledge Tracing (BKT) within adaptive learning systems. Using the ASSISTments 2017 dataset, a manually parameterised BKT baseline reached a Test AUC of 0.5715, while SAKT achieved a mean Test AUC of 0.669±0.001 over three runs. The resulting 0.097 AUC gap suggests attention-based models better capture sequential learner behaviour. Results support integrating transformer-based components in educational platforms. Future work should include comprehensive hyperparameter optimisation, significance testing, and cross-dataset validation.