The advent of big data and artificial intelligence (AI) technologies has catalyzed transformative changes across various sectors, including education. We propose two graph neural network-based models—MTGNN and DAGNN—for improving the prediction of student performance in educational settings. These models are evaluated using the Open University Learning Analytics Dataset (OULAD). MTGNN leverages multiple similarity metrics to build student relationship graphs that capture latent patterns, improving prediction outcomes. DAGNN enhances this approach by adding attention mechanisms and graph augmentation, leading to more accurate and robust predictions. Our experimental results show that both models outperform conventional baselines, especially in identifying at-risk students. This work demonstrates the potential of GNNs to transform educational analytics by modeling complex student data, enabling more effective, personalized interventions.