STAF-Net: A Spatio-Temporal Attention Fusion Network for Real-Time Student Engagement Detection in E-Learning Environments

The rapid growth of e-learning has made student engagement a critical factor in online education effectiveness.Traditional engagement measurement techniques, such as selfreports and interaction logs, are limited in capturing real-time behavioral cues.In this study, we propose a novel deep learning framework, the Spatio-Temporal Attention Fusion Network (STAF-Net), for classifying student engagement levels using body language captured from video.STAF-Net integrates spatial features extracted via a ResNet-based CNN with a temporal attention mechanism inspired by Transformer encoders, allowing the model to focus dynamically on informative frames across time.We evaluate STAF-Net on two benchmark datasets, DAiSEE and EmotiW, following a pipeline of frame sampling, normalization, and optimized training using cross-entropy loss and the Adam optimizer.Our model achieves 98.3% accuracy and 0.99 AUC on DAiSEE, and 98.9% accuracy with 0.998 AUC on EmotiW, outperforming conventional CNN, LSTM, and hybrid baselines.Qualitative analyses using Grad-CAM and attention visualizations demonstrate the model's interpretability.These results highlight STAF-Net's potential for enabling real-time, behavior-aware engagement detection in adaptive online learning platforms.