Intelligent Multi Model Ensemble for Engagement Prediction

For intelligent educational systems, the ability to monitor
and respond to student engagement in real time is essential for enhancing learning outcomes. However, existing models often lack adaptability and practical deployment potential, as they depend on single data modalities, rigid ensemble mechanisms, and post-session analysis. This study introduces an intelligent multimodal ensemble framework designed to address these challenges by predicting student engagement using predefined multimodal educational datasets that
include facial expressions, voice tone, physiological signals, and interaction logs. The proposed system leverages deep neural networks (CNNs for spatial and RNNs for temporal analysis) in combination with classical machine learning algorithms (SVMs and Decision Trees), integrated through an adaptive weighting mechanism that dynamically adjusts model contributions based on predictive confidence. Furthermore, explainable AI techniques, particularly SHAP, are incorporated to enhance transparency and interpretability.
Experimental evaluations across multiple educational contexts demonstrate the framework’s superior performance in terms of accuracy, generalization, and real-time efficiency. Unlike prior multimodal ensemble approaches, the proposed model uniquely combines adaptive confidence-based weighting and SHAP-driven interpretability, offering a balanced and deployable solution that bridges the gap between accuracy and explainability in real-world learning environments.