Automated UI usability evaluation and suggestions using Shneiderman's eight golden rules and deep learning

Usability evaluation of the user interface (UI) and user experience (UX) design, which directly influences system effectiveness, user satisfaction, and adoption, is a critical component. Although Shneiderman’s eight golden rules provide a well-established theoretical framework for usability assessment, their manual application is time-consuming, subjective, and difficult to scale within modern software development environments that involve rapid iterations and large numbers of interface screens. This study aimed to develop an Artificial Intelligence (AI)-based system for automated User Interface usability evaluation according to Shneiderman’s eight golden rules, with minimal reliance on human experts. A supervised multi-label deep learning framework was developed using annotated user interface screenshots. A dataset of over 4,000 screenshots from diverse application domains, including e-commerce, banking, and education, was curated and labelled for compliance with eight usability rules. Transfer learning was implemented using three different Convolutional Neural Networks (CNN) architectures: ResNet-18, ResNet-34, and ResNet-50, all trained under identical experimental settings with memory-aware optimizations. Among these three models, the ResNet-34 was found to be better performing in reducing the usability evaluation time from the standard 15-30 minutes per screen to seconds. Model performance was evaluated using subset accuracy, macro-averaged precision, recall, F1-score, and Hamming loss. Experimental results demonstrated that ResNet-34 eclipses the other architectures, achieving a subset accuracy of 85.71%, an F1-score of 0.3625, a precision of 0.3750, a recall of 0.3523, and a Hamming loss of 0.0179. The model achieved stable convergence at 26 epochs with good generalization ability, whereas the ResNet-18 achieved moderate performance and the ResNet-50 architecture faced major issues of overfitting and unstable training. The findings confirm the hypothesis that an automated usability evaluation based on convolutional neural networks, with an emphasis on established design standards, is indeed possible and effective. This proposed research delivers the potential for the inclusion of usability evaluation within modern software development processes, which would enable hybrid decision-making between humans and artificial intelligence to improve the quality of user interfaces.