Explainable AI Insights into Gender-Associated Facial and Upper-Body Patterns in Public Speaking: An Exploratory Study

Abstract

Using the 3MT_French dataset of three-minute thesis pitches, we investigate gender differences in non-verbal public-speaking expressivity through Facial Action Unit (AU) statistics and upper-body posture and movement cues. Four classifiers (RF, XGBoost, SVC, LR) trained under 10-fold stratified cross-validation reach best F1-scores of 0.72 for face-only and 0.62 for body-only, both with XGBoost. A SHAP-based feature ranking combined with FDR-corrected non-parametric tests reveals a two-axis structure: male speakers show a sustained baseline of activity in articulation-related AUs (notably Lip Pressor AU24) and a broader upper-body posture, while female speakers display greater dynamic facial expressivity and higher peak smile intensity (Lip Corner Puller AU12). While partially aligned with prior work, these findings extend known patterns to fine-grained AU statistics and reveal postural and dynamic gender-associated patterns in public speaking. Cross-corpus replication remains a key next step.