A Metric Hybrid Autoencoder Enabled Generative AI and Image Processing in IoT-Based Cloud Platforms for Analyzing Student Attention in Offline Classes

Accurately predicting student engagement is essential for improving teaching effectiveness and supporting students’ overall development. Computer vision has enabled automatic analysis of online classroom behaviors, but existing deep learning models struggle to capture subtle facial, posture, and gaze cues, and their performance is often limited by insufficient data. To overcome these challenges, we propose a Metric Hybrid Autoencoder (MHAE) model within a cloud platforms for engagement prediction. The model employs a variational autoencoder (VAE) for data augmentation, a convolutional autoencoder (CAE) for representation learning, and a metric learning module to enhance feature discrimination. This joint strategy improves the model’s ability to distinguish engagement states while ensuring robustness and generalization. Experiments on a real-world dataset demonstrate that MHAE outperforms state-of-the-art baselines, achieving 97.5% accuracy, 97.0% precision, and 97.1% F1-score. These results highlight the practical potential of MHAE to deliver reliable feedback and decision support in online education platforms.