End-to-end prior-guided graph attention network with differentiable structural priors for interpretable mobile interface layout scoring

Automated assessment of mobile-interface aesthetics remains difficult because professional judgments depend on spatial relations among components, whereas image-only models mainly summarize global appearance. This study develops an interpretable end-to-end scorer for static mobile screenshots. The proposed prior-guided graph attention network combines a YOLOv10 component detector, a ResNet-50 visual stream, and a graph stream whose attention is modulated by differentiable priors for proximity, alignment, white space, and touchtarget accessibility. Comparisons with convolutional, graph, and transformer baselines were conducted on app-level splits of the MUD and CLAY datasets. On MUD, the method obtained a weighted root mean square error of 0.387 and a Spearman rank correlation of 0.856, improving the strongest convolutional baseline by 14.3%. Ablation, detector-sensitivity, cross-dataset, and expert blind-test analyses showed that the structural priors improve accuracy, transferability, and diagnostic agreement while retaining end-to-end latency of 35 milliseconds per screen.