Metacognitive skills driven knowledge tracing

Knowledge Tracing (KT) monitors students’ knowledge state and predicts their learning performance. Existing KT methods primarily incorporate cognitive skills, e.g. concept mastery, and question-solving to characterize the student learning process. However, high-order metacognitive skills, which control cognitive skills, have not been fully considered. This paper proposes the Metacognitive Skills-driven Knowledge Tracing (MSKT) method and models the impact of metacognitive skills on the student's learning process. It extracts individualized metacognitive skills from learning behaviors and applies information bottleneck and self-attention fusion modules to align and fuse cognitive and metacognitive features. Comprehensive experiments were complemented on two real-world datasets, i.e. Ednet-KT4 and private dataset, along with baseline methods including machine learning-based, deep learning-based, and large language model-based. The results demonstrate that MSKT outperforms existing KT methods and achieves state-of-the-art performance. Visualization experiments show that MSKT further enhances the modeling of the student's learning process. The ablation study also illustrates that the introduced modules are reasonable and effective. These findings offer practical implications for the design of intelligent tutoring systems, suggesting that the integration of metacognitive skills could lead to more personalized and adaptive learning experiences.