Predicting academic success with graph neural networks
Abstract
Robust models for predicting student success are crucial for enhancing student engagement, academic achievement, and overall educational outcomes. This study investigates the application of three Graph Neural Network (GNN) architectures for predicting student success and explores how these predictions can support personalized course recommendation systems in higher education. The models are trained on a comprehensive multi-faculty dataset incorporating academic records, course enrollments, and performance indicators. A novel graph-based representation, termed the Academic Trajectory Graph (ATG), is proposed to capture the intricate relationships among students, courses, and performance outcomes. Experimental results demonstrate that GNN-based models achieve strong predictive performance, with the Graph Convolutional Network (GCN) outperforming other baselines. However, the results also reveal variability across faculties, reflecting the influence of faculty-specific data characteristics and underlying biases. These findings underscore both the potential and the challenges of applying GNNs in educational settings. Future work will focus on improving scalability, mitigating faculty-specific biases, and extending the framework to dynamic graphs and cross-institutional validation to ensure robust, equitable, and generalizable student success prediction models.