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

In real-world online education environments, automated analysis of student learning engagement faces fundamental challenges. While existing models score well on specific test sets, they are often computationally sluggish and poorly adaptable when confronted with lengthy classroom videos and highly variable student behavioral habits. The core of the problem is that traditional models are prone to taking "shortcuts": mistaking students' innate habits for evidence of disengagement. Meanwhile, models like Transformers are constrained by their prohibitive computational costs. To solve these problems, this paper proposes a new framework, i.e., MC-MIFA, which combines efficient sequence modeling with causal reasoning methods. Specifically, in the feature extraction part, we designed a novel Interleaved MambaVision backbone. Through hierarchical processing and spatio-temporal compression, this network effectively captures long-range global information in videos while maintaining near-linear computational complexity. In the decision logic part, this framework does not stop at simple data fitting but constructs engagement analysis as a causally-motivated structured model. By introducing independence constraints and feature disentanglement techniques, we forcibly separate the core content representing engagement from the bias representing personal habits in the high-dimensional feature space. Then, using a counterfactual dynamic fusion mechanism, we intervene at the moment the model makes a decision, forcing the question: "If we disregard this student's specific sitting habit, is their true thinking state still focused?" Experimental results on multiple challenging datasets, including DAiSEE, UBFC-Engagement, and SEED-IV, prove that MC-MIFA is not only highly competitive in accuracy and mean absolute error but also shows minimal performance degradation when facing entirely new students.