Modeling Teacher Nonverbal Immediacy: Refinement and Analysis

Abstract

This paper presents a refined computational approach for assessing teacher nonverbal immediacy (NVI) from classroom video recordings. Building on a previously published baseline model, we re-examined the original ten-feature representation to evaluate whether focused, theory-driven refinements could improve model performance without altering the preprocessing pipeline. The revised feature set includes distance variability, face visibility, and a merged negative affect channel, alongside existing measures of gesture intensity, perceived proximity, and facial expressions. Models were trained and validated on a dataset of 403 annotated 30-second video segments from German secondary school classrooms. We evaluated four classical regressors—linear regression, support vector regression, random forest, and extra trees—alongside the original multilayer perceptron (MLP). The refined MLP improved the correlation with human ratings from 0.44 to 0.49 and showed the most consistent prediction errors across teachers. Feature sensitivity analysis confirmed that the model’s predictions aligned with theoretical expectations: greater proximity, facial visibility, and expressive gestures were associated with higher immediacy. These results highlight the value of integrating behavioral theory into model design and support the use of automated tools for scalable, interpretable analysis of nonverbal communication in classroom settings.