Online education, especially Massive Open Online Courses (MOOCs), has seen higher dropout rates compared to traditional classrooms. Student Dropout Prediction (SDP) helps identify at-risk students by analyzing their online behavior, allowing for timely interventions to improve student retention and course completion rates. While traditional methods use feature vectors to capture student activity, recent graph-based algorithms have shown promise by effectively revealing relationships between entities, making them a compelling approach for dropout prediction. This paper introduces Student Dropout Prediction using Graph Attention and k-nearest-neighbors Link Augmentation (SDP-GAT-KLA), a novel approach that leverages Graph Neural Networks (GNNs) to generate student and course representations from graph topological data, rather than depending solely on traditional feature vectors. Using the KDD 2015 dataset, the proposed method first transforms the data into a bipartite graph, where students and courses serve as nodes and enrollment relationships as edges. We then apply a GNN model to extract meaningful representations from the bipartite student-course graph, incorporating both structural node features and activity-based edge features. These representations serve as inputs for deep neural networks to predict student dropout. Furthermore, to mitigate the impact of limited interactions caused by insufficient course data, which significantly affects prediction accuracy, we utilized a data augmentation strategy called K-Nearest-Neighbors Link Augmentation in dropout prediction area to enhance the robustness of the predictive model. The results of experiments conducted on the KDD 2015 dataset demonstrate the strong performance of our proposed method compared to baseline approaches.
