Combining a causal effect criterion for evaluation of facial attractiveness models

Attractiveness is an important attribute of a face that can be understood, predicted, and manipulated. Previous studies on facial attractiveness modeling focused on only prediction power. However, a model with high prediction accuracy may fail in attractiveness manipulation problems. In this paper, we add a causal effect criterion for model selection, which can be measured by imposing interventions according to the models and examining the change of attractiveness. We built a new database containing face images with diverse attractiveness and corresponding human ratings. Several manifold embedding and statistical regression methods are performed for attractiveness modeling. We compare different models under the two criteria and find that the performance of facial attractiveness manipulation depends on the causal effect of the model; feature normalization is a crucial step, without which the model will have small causal effect and fail in attractiveness manipulation; LDA manifold is better than PCA and LPP manifolds in attractiveness prediction; the linear model and the nonlinear SVR model are very similar, implying that facial attractiveness is quite directional. The selected model can be interpreted by common sense and works well for both attractiveness prediction and manipulation problems.