Multimodal feature importance ranking improves classroom teaching quality evaluation in human and artificial intelligence collaborative education

Abstract

Classroom Teaching Quality Evaluation (CTQE) constitutes a core component of human-AI collaborative education systems; however, existing methods exhibit significant deficiencies in the systematic importance ranking of multimodal features. This paper proposes the Multimodal Feature Importance Ranking via Multi-model Ensemble framework (MFIR-ME), which integrates five categories of modal features—acoustic, linguistic, visual, temporal, and behavioral—and incorporates a Cross-Modal Attention Fusion module (CMAF) to achieve dynamically weighted adaptive cross-modal information integration. Four complementary ranking methods, namely SHAP value analysis, Random Forest feature importance, XGBoost feature weights, and LIME local explanation, are further integrated through a Borda Count voting mechanism to generate robust consensus feature importance rankings. The framework instantiates human-AI collaboration through a three-stage feedback loop—AI-driven multimodal perception, teacher-facing explainability dashboard, and collaborative model refinement via teacher annotation—thereby preserving teacher agency while leveraging AI’s computational scalability. Systematic evaluation is conducted on a publicly available quantitative teaching evaluation dataset comprising 236 class sessions (11,328 min) across three subject types and six classes [1]. Teaching quality scores are evaluated under both a regression setting and a classification setting in which continuous scores are discretized into three ordinal tiers (High, Medium, and Low), enabling comprehensive assessment of continuous scoring precision alongside categorical discrimination ability. Results demonstrate that MFIR-ME achieves 88.6% accuracy in teaching quality prediction, significantly outperforming six baseline methods by 4.5%–−17.3%, while reducing MAE and RMSE by 22.5% and 19.8%, respectively. Linguistic semantic coherence, student engagement index, and teacher speech clarity are identified as the three most critical features influencing classroom teaching quality, exhibiting high cross-method consistency. Ablation experiments and robustness analyses further validate the modular effectiveness and noise robustness of the framework. This research provides the field of educational artificial intelligence with a multimodal teaching quality analysis tool that combines high-accuracy prediction with strong interpretability, offering data-driven and actionable guidance for teacher professional development.