Automated classification of demographics from face images: A tutorial and validation

Abstract

Examining disparities in social outcomes as a function of gender, age, or race has a long tradition in psychology and other social sciences. With an increasing availability of large naturalistic data sets, researchers are afforded the opportunity to study the effects of demographic characteristics with real‐world data and high statistical power. However, since traditional studies rely on human raters to asses demographic characteristics, limits in participant pools can hinder researchers from analyzing large data sets. Automated procedures offer a new solution to the classification of face images. Here, we present a tutorial on how to use two face classification algorithms, Face++ and Kairos. We also test and compare their accuracy under varying conditions and provide practical recommendations for their use. Drawing on two face databases (n = 2,805 images), we find that classification accuracy is (a) relatively high, with Kairos generally outperforming Face++ (b) similar for standardized and more variable images, and (c) dependent on target demographics. For example, accuracy was lower for Hispanic and Asian (vs. Black and White) targets. In sum, we propose that automated face classification can be a useful tool for researchers interested in studying the effects of demographic characteristics in large naturalistic data sets.

Racial Disparities in the Sharing Economy: Evidence from More Than 100,000 Airbnb Hosts across 14 Countries

“…Moreover, race classification algorithms rely on perceptual cues that are easily detectable and discriminate between different racial groups (e.g., skin color). Previous studies found relatively high levels of accuracy in race classification, especially for Black and White targets (Jaeger et al, 2020;Rhue & Clark, 2016). However, accuracy may be lower for racial groups that are more perceptually ambiguous (i.e., not characterized by unique and easily detectable facial characteristics).…”

“…The race of hosts is usually classified by participants, which means that sample sizes are constrained by the size of participant pools or research budgets. In the present studies, we circumvent this problem by relying on an algorithm to code hosts' race based on their profile photo (as described in Jaeger et al, 2020). This allowed us to analyze a substantially larger sample of listings (N = 108,798), which should enable us to estimate the racial price gap with more precision.…”

“…Our analysis focused on White, Black, and Asian hosts, as these categories can be classified with high levels of accuracy (Jaeger et al, 2020).…”

Presentation in self-posted facial images can expose sexual orientation: Implications for research and privacy.

“…Next, I cleaned and preprocessed all facial images with the help of the FacePlusPlus (Face++) application programming interface (API), which is a facial recognition software widely used in facial research (Kosinski, 2021; Wang et al, 2019; Wang & Kosinski, 2018), verified to be accurate at extracting facial information from images (Jaeger et al, 2020). Four sets of information were extracted: The number of faces in each facial image, facial landmarks, facial attributes, and facial bounding boxes.…”

Presentation in Self-Posted Facial Images Can Expose Sexual Orientation: Implications for Research and Privacy

“…Next, I cleaned and preprocessed all facial images with the help of the Face++ API, which is a facial recognition software widely used in facial research (Kosinski, 2021;Wang et al, 2019;Wang & Kosinski, 2018), verified to be accurate at extracting facial information from images (Jaeger et al, 2020). Four sets of information were extracted: the number of faces in each facial image, facial landmarks, facial attributes, and facial bounding boxes.…”