Taking Politics at Face Value: How Features Expose Ideology
Loading...
Authors
Copp, Christopher
Issue Date
2022
Type
Dissertation
Language
Keywords
Computer Vision , Facial Morphology , Gun Control , Immigration Policy , Neural Networks , Physiognomy
Alternative Title
Abstract
Previous studies using computer vision neural networks to analyze facial images have uncovered patterns in the feature extracted output that are indicative of individual dispositions. For example, Wang and Kosinski (2018) were able to predict the sexual orientation of a target from his or her facial image with surprising accuracy, while Kosinski (2021) was able to do the same in regards to political orientation. These studies suggest that computer vision neural networks can be used to classify people into categories using only their facial images.However, there is some ambiguity in regards to the degree to which these features extracted from facial images incorporate facial morphology when used to make predictions. Critics have suggested that a subject’s transient facial features, such as using makeup, having a tan, donning a beard, or wearing glasses, might be subtly indicative of group belonging (Agüera y Arcas et al., 2018). Further, previous research in this domain has found that accurate image categorization can occur without utilizing facial morphology at all, instead relying upon image brightness, color dominance, or the background of the image to make successful classifications (Leuner, 2019; Wang, 2022).
This dissertation seeks to bring some clarity to this domain. Using an application programming interface (API) for the popular social networking site Twitter, a sample of nearly a quarter million images of ideological organization followers was created. These images were followers of organizations supportive of, or oppositional to, the polarizing political issues of gun control and immigration. Through a series of strong comparisons, this research tests for the influence of facial morphology in image categorization. Facial images were converted into point and mesh coordinate representations of the subjects’ faces, thus eliminating the influence of transient facial features. Images were able to be classified using facial morphology alone at rates well above chance (64% accuracy across all models utilizing only facial points, 62% using facial mesh). These results provide the strongest evidence to date that images can be categorized into social categories by their facial morphology alone.