Abstract

Knowledge Tracing (KT) is a fundamental task in personalized learning, aiming to estimate a learner’s mastery level over time based on sequential exercise response interactions. Although deep knowledge tracing models have demonstrated superior predictive performance, their black-box nature severely hampers practical pedagogical adoption and actionable student feedback. In this paper, we explore an interpretable deep knowledge tracing framework by integrating Shapley Additive Explanations (SHAP) with Large Language Models (LLMs). SHAP quantifies the contribution of historical exercise interactions and knowledge concepts toward predicted performance, while LLMs synthesize these feature attribution scores into coherent, human-readable pedagogical explanations. Experimental evaluations demonstrate that this integration enhances transparency, facilitates diagnostic feedback, and maintains high predictive fidelity across benchmark educational datasets.