Student Learning Outcome Prediction via Sheaflet-Based Graph Learning and LLM

Accurately modeling the interactions between students and learning content is a central challenge in achieving personalized and adaptive learning in online education. However, existing methods often struggle to simultaneously capture the multi-scale structural dependencies and the rich semantic information embedded in educational materials. To bridge this gap, we propose EduSheaf—a unified framework that integrates large language models (LLMs) with a sheaflet-based signed graph neural network. Specifically, LLMs are employed to extract fine-grained semantic embeddings from multiple-choice questions (MCQs), thereby enriching graph representations with contextual knowledge. A signed graph is then constructed to encode student–MCQ interactions, where correct and incorrect responses are represented as positive and negative edges. On top of this, a novel sheaflet-based signed graph neural network performs multi-frequency learning through low-pass and high-pass filters, enabling the joint modeling of global consensus and local variations, while sheaf structures enforce edge-level consistency. Extensive experiments on multiple real-world educational datasets demonstrate that EduSheaf consistently outperforms state-of-the-art baselines, including both semantic-enhanced and signed graph models, in terms of prediction accuracy and robustness. Ablation studies further reveal the complementary roles of semantic embeddings and multi-frequency graph filters.