Enhancing Student Performance Prediction on Learnersourced Questions with SGNN-LLM Synergy

Learnersourcing offers great potential for scalable education through studentcontent creation. However, predicting student performance on learnersourcedquestions, which is essential for personalizing the learning experience, ischallenging due to the inherent noise in student-generated data. Moreover,while conventional graph-based methods can capture the complex network ofstudent and question interactions, they often fall short under cold startconditions where limited student engagement with questions yields sparse data.To address both challenges, we introduce an innovative strategy that synergizesthe potential of integrating Signed Graph Neural Networks (SGNNs) and LargeLanguage Model (LLM) embeddings. Our methodology employs a signed bipartitegraph to comprehensively model student answers, complemented by a contrastivelearning framework that enhances noise resilience. Furthermore, LLM'scontribution lies in generating foundational question embeddings, provingespecially advantageous in addressing cold start scenarios characterized bylimited graph data. Validation across five real-world datasets sourced from thePeerWise platform underscores our approach's effectiveness. Our methodoutperforms baselines, showcasing enhanced predictive accuracy and robustness.