Learning Causal and Sequential Patterns in Student Knowledge Tracing via Transformer-Augmented Bayesian Networks
Rachel Kim · Jisuanji shenghuojia. · 2025
Student knowledge tracing requires sophisticated modeling approaches that can capture both the causal relationships between learning concepts and the sequential patterns of knowledge acquisition over time. Traditional knowledge tracing methods struggle to simultaneously model causal dependencies and temporal learning dynamics while maintaining interpretability for educational practitioners and maintaining computational efficiency for real-time applications. The challenge lies in developing frameworks that can learn complex causal structures from educational data while effectively capturing the sequential nature of learning processes and providing actionable insights for personalized education. This study proposes a novel Transformer-Augmented Bayesian Network (TABN) framework that integrates transformer architectures with probabilistic graphical models to enable comprehensive modeling of causal and sequential patterns in student knowledge tracing. The framework employs transformer networks to capture long-range sequential dependencies in learning trajectories while utilizing Bayesian networks to model causal relationships between knowledge concepts. The integrated approach enables joint learning of causal structures and sequential patterns through end-to-end optimization while maintaining probabilistic interpretability essential for educational applications. Experimental evaluation using large-scale educational datasets demonstrates that the proposed framework achieves 41% improvement in knowledge tracing accuracy compared to traditional methods. The TABN approach results in 36% better prediction of learning outcomes and 44% improvement in causal relationship discovery between knowledge concepts. The framework successfully combines the sequential modeling capabilities of transformers with the causal reasoning advantages of Bayesian networks, resulting in 32% better interpretability scores and 28% improvement in educational decision support compared to existing knowledge tracing approaches.