Dropout prediction using Graph Neural Network

Abstract

Dropout prediction leverages educational big data from CBT and other assessments. It forecasts the likelihood that a student will fail to complete their course. Student dropout impacts academic and career trajectories and negatively affects mental well-being. Therefore, accurate and interpretable prediction models are crucial for identifying at-risk students and providing support. This study proposes a dropout prediction using hierarchical graph neural networks. We construct an item graph where nodes represent assessment items in the same field. Then, we apply DiffPool, a graph pooling method, to generate graph embeddings, which are assigned as feature vectors to the field graph nodes. We apply DiffPool again to the field graph to compute dropout prediction scores. This hierarchical approach captures the structural relationships among input features and enhances interpretability through graph visualization and DiffPool weight analysis. Experimental results show the proposed method outperforms conventional approaches in accuracy. We also analyze the optimized graph structure to explore inter-field relationships and identify key academic domains influencing dropout prediction.