Data augmented hybrid GCN transformer for student engagement recognition in E-learning

Student engagement plays a critical role in effective educational activities within e-learning environments. However, automatic engagement recognition from webcam videos remains challenging due to imbalanced affective data distributions and the subtle nature of facial expressions. This paper proposes a data-augmented hybrid framework that integrates graph-based geometric modeling with transformer-based temporal learning to recognize student engagement levels from facial activities. A variational autoencoder is employed to generate semantically consistent synthetic facial samples, alleviating performance degradation caused by severe class imbalance in facial engagement datasets. The proposed graph-based model explicitly captures multi-scale geometric relationships among facial landmarks and action units, while the transformer architecture enriches engagement representations by modeling long-range temporal correlations in facial dynamics. Experiments conducted on the DAiSEE benchmark demonstrate that the proposed framework achieves an F1-score of 72.89% and an accuracy of 71.25%, outperforming state-of-the-art temporal convolutional, recurrent, and transformer-based engagement recognition methods. Ablation studies further confirm the complementary contributions of generative data augmentation and topology-aware geometric modeling, yielding performance improvements with negligible computational overhead. The proposed approach provides a robust and reliable solution for student engagement monitoring in real-world e-learning scenarios.