Deep Attention Knowledge Tracing Based on Temporal Kolmogorov-Arnold Networks

As a fundamental task in adaptive learning systems, Knowledge Tracing (KT) involves predicting how well a student will perform based on their previous answering behavior. Deep knowledge Tracing models in earlier studies are often built upon RNN architectures to represent the evolving knowledge states of learners. However, these models have limitations in capturing long-term dependencies and tend to fix the activation function to handle time steps without considering the time-decay effect of students’ responses. Therefore, it is difficult to accurately characterize the complex behavior patterns of students and the dynamic evolution of their mastery of knowledge. To address these issues, this paper proposes a Deep Attention Knowledge Tracing model based on the Temporal Kolmogorov-Arnold Network (DATKAN-KT). The model first employs an attention mechanism to explore the relationships between the importance of historical student responses, while incorporating the time decay effect of their answers. Additionally, to learn more complex dependencies in long sequences of student responses, the model employs the Temporal Kolmogorov-Arnold Network to capture students’ evolving knowledge representations from long answering sequences, with the goal of predicting future response outcomes. To evaluate the proposed model’s effectiveness, extensive experiments were conducted on multiple publicly available KT datasets. The evaluation shows that DATKAN-KT performs more effectively than existing KT models in four public datasets and remains robust when handling extended response sequences, underscoring its ability to capture the evolving knowledge states of students.