Predicting Student Programming Performance Based on LLM Feature Enhancement and Heterogeneous Graph Neural Networks

• In Event: Computer Science Education and Computational Thinking

Abstract

This study proposes a novel approach to predict student programming performance using Large Language Model (LLM) feature enhancement and heterogeneous graph neural networks. This method models student programming using submission nodes, bridging student attributes and problem features through code content. We frame performance prediction as a node classification task within graph neural networks and introduce an LLM-to-LM (traditional Language Models) strategy to enhance feature representations. Experiments on an authentic online programming platform dataset containing over 160,000 submissions demonstrate significant improvements in prediction performance compared to traditional methods and other graph neural network baselines. Ablation studies validate the effectiveness of our approach. This study enables early identification of at-risk students and supports the development of personalized learning paths.

Authors

• Shan Li, Lehigh University

• Hao Wu, Guangzhou University

• Guozhu Ding, Guangzhou University