MC-MIFA: a causal-aware hybrid state space framework for robust multimodal student engagement analysis

Area of Science:

• Artificial Intelligence
• Educational Technology
• Computer Vision

Background:

• Automated student engagement analysis in online education is challenging due to computational demands and variability in student behavior.
• Existing models often misinterpret personal habits as disengagement and struggle with long videos.
• Transformer models face high computational costs, limiting their practical application.

Purpose of the Study:

• To propose a novel framework, MC-MIFA, for robust and efficient automated analysis of student learning engagement.
• To overcome the limitations of existing models in terms of computational efficiency and adaptability to diverse student behaviors.
• To accurately differentiate genuine engagement from ingrained personal habits in online learning environments.

Main Methods:

• Developed a novel Interleaved MambaVision backbone for efficient feature extraction with near-linear computational complexity.
• Employed causal reasoning, independence constraints, and feature disentanglement to separate engagement signals from behavioral biases.
• Utilized a counterfactual dynamic fusion mechanism to assess engagement independent of personal habits.

Main Results:

• MC-MIFA demonstrates competitive accuracy and minimal mean absolute error on benchmark datasets (DAiSEE, UBFC-Engagement, SEED-IV).
• The framework exhibits robust performance, with minimal degradation when encountering unseen students.
• Effectively disentangles engagement from personal habits, addressing a key limitation of prior methods.

Conclusions:

• MC-MIFA offers an efficient and adaptable solution for automated student engagement analysis in real-world online education.
• The causal reasoning approach significantly improves the reliability and fairness of engagement detection.
• This framework has the potential to enhance personalized learning experiences by accurately understanding student focus.