Adaptive Heterogeneous Multi-Model Ensemble Framework for Modeling and Enhancing Student Engagement Using Multimodal Educational Data

This study introduces an Adaptive Heterogeneous Multi-Model Ensemble
Framework to accurately predict and model student engagement using multimodal
educational data. By analyzing facial expressions, voice intonation, physiological signals,
and interaction patterns, the framework adapts swiftly to dynamic changes in learner
behavior for strong predictive performance. The ensemble employs an adaptive weighting
mechanism, adjusting contributions from different models in real time to enhance robustness
across online, hybrid, and in-person learning environments. Unlike high-cost physiological
sensing approaches, this method leverages non-invasive modalities such as webcam-based
engagement detection and digital interaction analytics for a more accessible solution. To
address privacy and ethical concerns, privacy-preserving techniques like federated learning
and differential privacy are integrated, ensuring secure data processing and compliance. The
framework is benchmarked against traditional single-model methods, showing significant
improvements in accuracy, adaptability, and interpretability. The approach provides real-time,
explainable insights, enabling educators to deliver personalized interventions and strategies.
The framework fosters the development of AI-driven educational systems that promote
interactive and adaptive learning experiences. Emphasis is placed on maintaining student
privacy while empowering data-driven innovations in education