Computational models of facial attractiveness judgments.

We designed two computational models to replicate human facial attractiveness ratings. The primary model used partial least squares (PLS) to identify image factors associated with facial attractiveness from facial images and attractiveness ratings of those images. For comparison we also made a model similar to previous models of facial attractiveness, in that it used manually derived measurements between features as inputs, though we took the additional step of dimensionality reduction via principal component analysis (PCA) and weighting of PCA dimensions via a perceptron. Strikingly, both models produced estimates of facial attractiveness that were indistinguishable from human ratings. Because PLS extracts a small number of image factors from the facial images that covary with attractiveness ratings of the images, it is possible to determine the information used by the model. The image factors that the model discovered correspond to two of the main contemporary hypotheses of averageness judgments: facial attractiveness and sexual dimorphism. In contrast, facial symmetry was not important to the model, and an explicit feature-based measurement of symmetry was not correlated with human judgments of facial attractiveness. This provides novel evidence for the importance of averageness and sexual dimorphism, but not symmetry, in human judgments of facial attractiveness.