Internal and Observable Correlates of Perceived Dominance: A Multimodal Study of Audio and Physiology in Small-Group Interaction

Abstract

Perceived dominance is an interactional attribute that emerges from observable behaviour during small-group interaction and is systematically perceived by group members. Prior computational work has established audio and conversational-floor features as robust predictors of perceived dominance in meeting corpora, but whether physiological signals indexing autonomic arousal add predictive value beyond the observable behavioural stream remains an open question. We investigate this question using a corpus of ten four-person groups that completed four collaborative tasks. Each participant wore a close-talk microphone and a wrist-worn electrodermal and photoplethysmographic sensor. Every group member rated the dominance of each participant at the end of each task in a round-robin design. An intraclass correlation analysis of these ratings shows that the three-peer mean is moderately reliable, with the greatest consensus observed in the negotiation task, supporting its use as the prediction target. Following an ablation design, we compare ridge and elastic-net regression models trained on conversational-floor-only, full-audio, physiology-only, and observable-plus-physiology feature sets under leave-one-group-out cross-validation. Collinearity and redundancy diagnostics justify the use of penalised rather than ordinary regression. Conversational-floor behaviour predicts the averaged peer rating moderately well, whereas acoustic features alone provide little generalisable signal. Across estimators and feature representations, paired group-held-out sign-flip tests show that wrist physiology adds no reliable predictive value beyond the measured speech and conversational features. Relative to the strongest behavioural baseline, its increment is consistently negative. Stability selection identifies the behavioural signals that robustly convey perceived dominance.