Main content
Tracking historical changes in perceived trustworthiness using machine learning analyses of facial cues in paintings
Date created: | Last Updated:
: DOI | ARK
Creating DOI. Please wait...
Category: Project
Description: Open access data and scripts. The random forest model and the optimization script can be downloaded upon request for research and education purposes only. Disclaimer: The present algorithm was built to mimic how human raters would perceive trustworthiness of faces, based on a set of facial muscles activations. This algorithm was constructed based on a perception model, created from evaluations provided by Princeton University students on Caucasian faces, it thus incorporates the biases present in this population when providing first impressions ratings on these faces. The scores computed by the algorithm approximate human raters’ subjective perceptions of trustworthiness of the processed images based on the pose taken by the individual. It does not provide information about individuals’ actual character, including their trustworthiness. Therefore, it cannot be used to infer the actual level of trustworthiness of a person. Since the algorithm is highly sensitive to factors such as lighting and pose, it cannot be used to infer the general perception of an individual outside the precise portrait processed by the algorithm. This tool can be used to estimate how human raters would judge specific images of faces. It is interesting to assess the perception of portraits, which include cultural or historical elements (such as specific outfits, haircut, image quality or painting style) that contemporary raters may incorporate in their evaluations and which may not make evaluations comparable between time periods. It can be used to conduct historical and cross-national trend analyses based on large databases of portraits and pictures. However, as the estimation of muscle activation on which this tool is based on can be influenced by sitters’ fixed facial features, this tool cannot be used to infer the specific social intentions of a person, even on a specific picture. Importantly, this algorithm has no significant benefit compared to collecting actual human ratings for the evaluations of contemporary pictures as well as of images from which it is possible to remove contextual or historical information.