Interpretable knowledge tracing via fine-grained multi-feature attribution

Abstract

With the growth of massive educational data and the rapid advancement of artificial intelligence technologies, knowledge tracing has become increasingly important for assessing students’ knowledge states. Existing deep learning-based knowledge tracing models have achieved increasingly high predictive accuracy. However, they fail to capture significant features with explicit educational significance, which limits educators’ understanding, trust, and practical use of the diagnostic results. In this paper, we propose a Fine-Grained Multi-Feature Attribution Interpretable Knowledge Tracing model (MFA-IKT for short). It integrates educational theories with students’ learning behavior pattern, modeling fine-grained features of questions in terms of difficulty and discrimination and capturing the multidimensional dynamic features of students on knowledge mastery and ability profile. A Tree-Augmented Naive Bayes structure is adopted to construct the dependencies between the evidence features and the prediction outcomes. Experiments on five real-world datasets show that our model outperforms all baselines, including deep learning-based models, achieving average improvements of 9.28% in AUC and 9.99% in RMSE. Further analysis reveals that question-side features have a greater impact than student-side features. Among the fine-grained question features, discriminative features significantly enhance the model’s predictive performance. This study, through modeling interpretable features and attributing prediction outcomes, presents an explainable intelligent tutoring framework for personalized education, comprising “learning outcome prediction $\to$ feature attribution $\to$ instructional intervention suggestions”.