PREDICTING VISUAL AESTHETICS DISTRIBUTIONS OF MOBILE GUIS USING DEEP LEARNING

Visual aesthetics is a fundamental yet complex attribute of graphical user interfaces (GUIs), often represented by a single score.However, such simplification may overlook the variety of responses that GUIs elicit from different users, even when they share similar backgrounds.To address this limitation, we propose representing visual aesthetics as a distribution of ratings obtained from individual raters.From this distribution, the median can be interpreted as the degree of visual aesthetics, while the average absolute deviation reflects the level of disagreement among raters.Additionally, the histogram provides a more comprehensive visual analysis.In this work, we present a supervised deep learning approach to predict these distributions and compare different model architectures when assessing GUIs created with App Inventor.Our best results were achieved with a ResNet50 model, which obtained a mean squared error (MSE) of .021.While the correlation between individual score predictions and labels was moderate, the correlation between medians was strong ( = .85).Furthermore, the predicted medians showed excellent agreement with a model trained to predict single-valued aesthetics ( = .97),and Bland-Altman analysis revealed 95% agreement with ground-truth labels.These findings demonstrate that it is feasible to automate the prediction of visual aesthetics in mobile GUIs not only as single values but also as full distributions of human ratings, providing richer and more reliable insights into user perception.