Pei-Xuan Luo¹ (), Denise Hsien Wu¹, Erik Chihhung Chang¹; ¹National Central University, Taiwan

Whether facial attractiveness is universal or subject to the viewers’ own experience and demographic characteristics has been a key focus of research. While some studies demonstrated universal preferences such as higher attractiveness ratings for female than male faces regardless of demographical categories, other studies showed higher consistency of the attractiveness ratings from raters on own-race than other-race faces. Deep learning models, particularly convolutional neural networks (CNNs), have shown remarkable accuracy in predicting human ratings of facial attractiveness. Hence, the current study aims to examine how CNNs make attractiveness judgments on cross-category faces that they were not specifically trained for, offering insights into human facial attractiveness judgments. We fine-tuned two CNN models, GoogLeNet and ResNeXt-50, with facial images and their attractiveness ratings from human participants in the SCUT-FBP5500 (N=5500) and a custom (N=346) dataset. Cares were taken to balance race (Caucasian/East Asian) and gender (male/female) of faces for training. Five-fold cross-validation assessed model performance via Pearson’s correlation coefficients between human ratings and CNN predictions. The GoogLeNet and ResNeXt-50 models trained with a specific demographic category achieved average correlations of 0.78 and 0.69, respectively. When applying these models to predict attractiveness of faces in the same or a different demographic category, both models showed higher prediction accuracy for within-category than cross-category faces, indicating a reliance on learned demographic-specific attributes or patterns. Overall, these results demonstrated that CNN models, trained to mimic human judgements, demonstrate higher sensitivity to within-category than between-category preferences when predicting facial attractiveness for novel faces. These findings corroborate the idea that demographic characteristics such as race and gender influence attractiveness judgments, aligning with the role of prior experience in shaping preferences.

Acknowledgements: This study was supported by grants provided by the National Science and Technology Council, Taiwan (NSTC 111-2410-H-008-060-MY3, NSTC 113-2410-H-008-073).