Abstract: Currently,academic performance prediction,as a core component of personalized educational support systems,has become a focal point of research in the field of educational data mining,playing a significant role in optimizing teaching decisions and guiding student development.However,traditional prediction methods struggle to effectively address the challenges posed by the complex correlations,temporal evolution,and group dependencies inherent in multi-source heterogeneous data within educational contexts,resulting in limitations in prediction accuracy and generalization capabilities.Graph Neural Networks(GNNs),leveraging their powerful relational modeling and representation learning abilities,provide a novel paradigm for addressing these challenges.Consequently,numerous researchers are dedicated to applying GNNs to academic performance prediction research.This paper presents a systematic review of current research efforts on GNN-based academic performance prediction tasks.Starting from the problem definition,it analyzes the core challenges of academic performance prediction.It then outlines the foundational knowledge and common models of GNNs.Subsequently,it categorizes and reviews the representative models and their application scenarios for academic performance prediction,including static feature modeling,combined static and dynamic feature modeling,and techniques empowered by emerging large model technologies.Building on this,the paper systematically summarizes and analyzes the evaluation-related datasets and metrics used for GNN-based academic performance prediction methods.Finally,it prospects future research directions from perspectives such as model scalability,interpretability,multimodal semantic information fusion,and dynamic graph pre-training.

Key words: Graph neural network, Academic performance prediction, Static features, Dynamic features, Large models, Educational data mining