Beauty is in the eye of your cohort: Structured individual differences allow predictions of individualized aesthetic ratings of images.

Abstract

In the last few years, there has been an increasing interest in computational models that are capable of predicting the aesthetic ratings of images based on objective image features. Given that aesthetic ratings vary across individuals, models that predict the average aesthetic ratings ignore the unique taste of an individual. In this paper, our goal is to better understand the individual differences in aesthetic ratings by investigating if individual differences follow structural rules or if taste is due to a random component of an individual's ratings. We address this question by using a collaborative filtering model that uses the similarities in ratings of a cohort of observers to predict individuals' ratings on a new set of images. Using Amazon Mechanical Turk, 299 online participants were instructed to rate how much they like a set of 50 art images. Using a subset of the images (40), we formed cohorts of individuals with similar ratings and used these cohorts to predict how each person would rate the remaining 10 images not included in the training set. The selected cohorts predicted individual ratings significantly better than random cohorts and outperformed predictions based on the mean image ratings. We also found that the optimal size was approximately 12 % of the sample size. These results imply that individual differences in fact have an underlying structure that is consistent across the cohort and are not random. Using personality scores and subject backgrounds, we also looked at the subject characteristics of the cohorts and found that the participants' art background was the only significant factor. Finally, we explored whether the cohorts used particular visual features in a consistent way. For our small set of features, we didn't find any evidence for this. These results provide important insights into the sources of individual differences in aesthetic preferences and highlight the potential for computational models to improve predictions of individual preferences by leveraging structured individual differences rather than relying solely on population averages.