Deployment Challenges of Bayesian Network–Based Learning Path Recommendation in Real Classroom Settings

Bayesian Network (BN)–based Learning Path Recommendation (LPR) systems are widely adopted in personalized education for modeling uncertainty and providing interpretable learner representations. However, existing studies predominantly evaluate these systems under controlled settings that assume balanced data, simplified curricula, and unconstrained resources. Consequently, limited empirical understanding exists regarding their performance in authentic classrooms. This study addresses this gap by examining the real-world deployment of a BN-based LPR system using naturally occurring classroom data. The system is evaluated using 426,004 quiz responses collected across 19 formative assessments, reflecting routine instructional practice. Two curricular configurations were examined: an independent-topic structure with no explicit dependencies, and a prerequisite-based structure where mastery depended on prior knowledge. Rather than focusing on predictive accuracy, the study investigates operational feasibility, including structural complexity and memory consumption. From a software-engineering perspective, the central problem addressed is the scalability and inference feasibility of BN-based LPR systems under curriculum dependencies and real-world resource constraints. Results show that BN inference remains stable and computationally tractable for independent topics. In contrast, prerequisite-based modeling substantially increases network density, memory requirements, and inference variability, in some cases causing deployment failures under standard resources. Sparse assessment coverage at higher cognitive levels further undermines recommendation stability. This work provides a deployment-oriented perspective that complements performance-centric research and emphasizes the need for deployment-sensitive evaluation beyond accuracy-focused assessment in real classrooms.