SymCART: a symbiotic cognitive-affective reinforcement transformer for optimizing educational interventions

Emotion and cognition analysis in educational interventions is crucial for enhancing personalized learning outcomes. However, existing models often encounter challenges in multimodal data integration and adaptive strategy optimization. To address these challenges, we propose the SymCART model, an intelligent educational intervention framework that integrates multimodal data fusion with deep reinforcement learning-based strategy optimization. SymCART dynamically adjusts teaching strategies through the collaborative operation of a multimodal perception encoder, dynamic cognitive-affective graph inference engine, and adaptive teaching strategy optimizer, thereby improving student learning outcomes. Experimental results demonstrate that SymCART achieves higher predictive accuracy and more effective strategy recommendations compared to traditional models, with statistically significant improvements in AUC, RMSE, weighted F1 for predictive tasks, and nDCG and ADR for policy/recommendation tasks across the IMPROVE, student learning behavior, and additional validation datasets. Ablation studies further confirm the essential contribution of each module, particularly regarding multimodal fusion and strategy optimization. The SymCART model provides robust support for personalized educational interventions and exhibits broad applicability for emotion and cognition analysis as well as adaptive learning strategies.