A simulation tournament for facial attractiveness to select stimuli.

In this project, we aim to evaluate whether we can use the "Elo framework" to aid in the selection of stimuli judgments of attractiveness. A common problem in attractiveness research is selecting stimuli. These stimuli are either targets or then used as a base stimulus for further transformations (morphing). Typically, this selection is achieved by either picking stimuli that are above a certain threshold (e.g., a certain percentile or a certain number of standard deviations away from the means or medians). However, these decisions are typically arbitrary and based on aggregate ratings (means / medians) rather than individual raters' choices. We propose to use methods from animal research: an Elo framework. This framework allows a more precise quantification of attractiveness. We aim to evaluate the usefulness of this Elo framework to study ratings of facial attractiveness in a widely used set. The purpose is threefold. First, we want to describe the individual scores for each face, i.e. an Elo score, which can be thought of as a dominant score. In this context, how much more attractive a face is compared to another face. Second, we want to describe the properties of this hierarchy, e.g. steepness, of this set of faces. Third, we want to evaluate how stable the scores and the hierarchy are when imposing different decision rules (e.g., imposing a rule that ratings must be opposite sex). To this end, we simulate tournaments of contests between the stimuli whereby the more attractive face 'wins'. We will then analyze these contests using the Elo framework. The added benefit of this framework is that for each stimulus we can generate an estimated 'win' probability (here: the likelihood that it will be found more attractive than another one). For example, an Elo difference between two stimuli of 200 points indicates that one stimulus is more attractive than the other one in 76% of the contests. One heuristic that is often used in chess is a difference of 400 points in Elo. This corresponds to a win probability of 90.9% or roughly a 10:1 odds ratio for the attractive stimulus. Phrased differently, the unattractive stimulus is expected to win 1 in 11 contests. We compare which stimuli would be chosen when applying an Elo heuristic, such as the 400-point difference rule, to generate attractive versus unattractive stimuli sets versus a different heuristic (Mean +/- SD rule) to define attractive versus unattractive stimuli sets. We will also compare to a "top 5 rule", picking the top-ranked stimuli when ordered by mean attractiveness ratings. (Conversely, we will also evaluate a "bottom 5" rule for The Elo system is further explained in the 'Illustration' folder of the associated project.