FRAME: feature-oriented retrieval and clustering of app screens

What happened on this date

At 19:39 UTC on 2026-08-09, arXiv received v1 of "Fusing UI Structure & Semantics for Feature-Oriented App Screen Retrieval & Clustering" (primary category cs.SE), with a comment that it was accepted at ICSME 2026. The authors note that UI programming tools often depend on screen retrieval and clustering, and that computing similarity by feature is hard because screens with the same function frequently look different. They propose FRAME, a multi-modal, neuro-symbolic screen embedding: it builds graph-based symbolic representations of UI components to capture key relationships and feature patterns across screens, then uses large vision-language models for visual and lexical encoding together with a UI-specific computational geometry algorithm for weighted embedding propagation. The abstract reports up to 13% higher MRR in search and up to 7.6 percentage points higher clustering accuracy than strong baselines across three benchmarks, with an ablation study supporting each component.

Why it matters for people protecting app logic

A boundary first: this paper aims to improve UI design and testing tools and does not study clone detection; what follows is our inference from its method. The UI is the part of an app that cannot be hidden, since every user can see and screenshot it. When a model can group screens that look different but do the same thing, mapping an app's complete feature flow becomes more automated, and that cuts both ways. Imitators can more easily line up and reproduce your flows, and you can apply similar ideas to spot re-skinned look-alikes in app stores. The resulting conclusion is plain: code obfuscation does not protect UI flows, and what should be hard to copy is your server-side logic, data and operational capability.

Concrete checks

First, list UI flows, copy and interactions as public information in your threat model and do not budget code-level protection for them. Second, make sure the key value does not live only in the UI layer: pricing, recommendation and risk rules should execute on the server. Third, keep dated design files, version screenshots and release records as evidence for later originality claims. Fourth, if you monitor for imitation apps, compare by feature flow and key screen structure rather than only icons and color schemes, which are the easiest things to change.

Your checklist

• Treat UI flows as public in your threat model
• Confirm pricing, recommendation and risk rules run server-side
• Keep dated design files and version screenshots
• Compare look-alikes by feature flow, not just icons and colors