| dc.description.abstract | Understanding student learning behavior is an important task in modern educational technology.
Knowledge Tracing (KT) can estimate students’ changing knowledge states from their previous
learning interactions and predict future performance. This study investigates the use of deep
learning-based Knowledge Tracing models to analyze student learning behavior using real
interaction data collected through an online quiz platform.
The dataset contains 6,318 interactions from 426 students covering 12 skills and 3 courses. The data
were cleaned and chronologically ordered before modeling. Two behavioral features, Time Gap and
Attempt Count, were introduced to examine their effect on prediction performance. Four models,
DKT, DKVMN, SAKT, and AKT-lite, were evaluated using four feature configurations with
Grouped 5-Fold Cross-Validation.
The results show that DKT with Time Gap and Attempt Count achieved the highest pooled ROC
AUC of 0.775. DKVMN achieved 0.684, AKT-lite 0.674, and SAKT 0.616. Attempt Count was the
most consistently useful engineered feature across the evaluated models. The study also explores
knowledge-state analysis, adaptive question sequencing, and learner grouping to demonstrate the
practical potential of Knowledge Tracing. | en_US |