Aesthetics Prediction of Websites using Federated Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Interface aesthetics is considered subjective in nature. Most existing approaches to aesthetic prediction generally rely on centralized data collection and use average ratings, often ignoring individual user choices. To address this, we propose a Federated Learning-based model. To construct the base model, we considered EfficientNet-B0, ResNet- 50, and Vision Transformer (ViT-B/16). To test these three models, we developed a dataset of 100 website interfaces (50 mobile and 50 desktop/laptop) and had 100 users rate them using the Latin Square Method. It was observed that EfficientNet-B0 outperformed the others. Consequently, we considered EfficientNet-B0 as the base model. In the empirical study, we found that users can be grouped into five types based on their gender and occupation. Accordingly, five clients were considered for our Federated Learning algorithm. The Federated Average (FedAvg) algorithm achieved a promising MSE of only 0.0534.