Vision transformer-based student engagement recognition in online economic law education

Student engagement recognition in online legal education faces significant challenges due to the absence of direct instructor-student interaction and real-time feedback mechanisms. This study proposes an Enhanced Vision Transformer (E-ViT) framework specifically designed for automated student engagement detection in online economic law courses. The framework introduces three key innovations: (1) multi-scale patch embedding that simultaneously captures coarsegrained body posture and fine-grained facial micro-expressions, (2) spatial-temporal attention mechanism with engagement-aware positional encoding for modeling dynamic patterns across video frames, and (3) hierarchical multitask learning architecture that jointly predicts cognitive, emotional, and behavioral engagement dimensions. Experimental evaluation on the Student Engagement Dataset demonstrates that E-ViT achieves 88.73% overall accuracy and 87.56% macro F1-score, outperforming standard Vision Transformer by 3.02% and traditional machine learning methods by over 17%. Comparative analysis against eight baseline methods including SVM, Random Forest, CNNs, and standard ViT validates the effectiveness of domain-specific architectural adaptations. The proposed framework provides interpretable multi-dimensional engagement predictions that enable educators to implement targeted pedagogical interventions in real-time. This research contributes to advancing AI-assisted online education by demonstrating how transformer-based architectures can be effectively adapted for specialized educational applications, paving the way for more effective and equitable remote legal education systems.