Abstract
First impressions of key social traits such as trustworthiness and competence are often based on rapid judgments of facial appearance, with substantial downstream consequences for individuals. A challenge remains to understand the specific face features that drive social perception. Here, we address this by showing that two major models of social perception (warmth/competence, Fiske et al., 2006; trustworthiness/dominance, Oosterhof & Todorov, 2008) are structured by a set of latent features that are shared across social traits, plus a set of trait-specific features that distinguish them. Specifically, we used a novel 3D face generator (Zhan, Garrod, van Rijsbergen, & Schyns, 2019), reverse correlation, social trait perception, and a data reduction technique to model these shared and unique features. Thirty participants (15 women, white, Western, 18-35 years) each viewed 2400 randomly generated 3D face identities and rated each on the four bipolar social trait dimensions (e.g., ‘very submissive’ to ‘very dominant’) in separate tasks. To identify the specific 3D face features that elicit these perceptions, we linearly regressed the 3D face information presented on each trial and the participant’s responses, producing 360 3D face models per trait. Next, to identify their features, we reduced all 3D face models with non-negative matrix factorization and mapped the resulting feature combinations that characterize each trait. Dominance and competence share an inwards change of the eyebrow region, also shared with cold and untrustworthy. Thus, a single feature can be shared across traits, including those thought to be unrelated with each other. Trustworthiness and warmth similarly share many features, predominantly around the mouth (i.e., upturned mouth corners). Our results reveal a compositionality of social trait perception, driven by shared 3D shape features plus unique accents, which have the generative capacity of designing digital social avatars and robots that convey first impressions of key social traits.