Multimodal student behavior analysis based on unsupervised learning

This study applies unsupervised learning to multimodal data for analyzing student behavior patterns, aiming to discover intrinsic behavioral structures without predetermined categories. The research integrates video, audio, digital interaction, and physiological signals across diverse educational settings. The framework employs self-supervised representation learning, multi-view clustering, and temporal pattern mining to analyze synchronized multimodal data streams from classroom, online, and laboratory environments. Five distinct behavioral clusters were identified, showing significant correlations with academic outcomes. The temporal stability of behavioral states emerged as a stronger predictor of achievement than frequency. The multimodal approach demonstrated superior performance compared to single-modality analyses in capturing behavioral transitions and detecting disengagement. Unsupervised multimodal analysis effectively reveals naturally occurring behavioral patterns adaptable across diverse educational contexts, establishing a methodological foundation beyond predefined categories. The approach enables earlier intervention for struggling students, improving upon traditional identification methods. These findings support the development of adaptive educational technologies that respond to students' behavioral states in real-time, enhancing personalized learning experiences.