Predicting Facial Impressions Using Guided Grad-CAM and GPT-4o

Facial impressions play a critical role in social interactions and, influence judgments of attractiveness, dominance, and trustworthiness. While traditional psychophysical experiments have investigate these impressions, computational approaches, particularly deep learning models, now offer a powerful means to uncover important facial features. This study employed the Visual Geometry Group 19(VGG19)-layer network deep learning model, integrated with explainable Artificial Intelligence (AI) techniques—Guided Gradient-weighted Class Activation Mapping (Guided Grad-CAM) and Generative Pre-training Transformer 4 omni (GPT-4o)— to predict and explain facial impressions. Guided Grad-CAM highlighted detailed facial regions relevant to specific impressions, whereas GPT-4o provided linguistic descriptions of these visualizations. The results demonstrated that key predictors included the eyes, surrounding eye region, and forehead for attractiveness; structural facial elements for dominance; and expressions for trustworthiness. This study offers new insights into the interpretability of facial impression predictions and proposes a novel approach to examining how specific facial features influence social judgments through deep learning and explainable AI techniques.