Abstract
Traditional app recommendation systems suffer from the "cold-start" problem, where high-quality new apps are overlooked due to a lack of historical user data like ratings or download counts. This paper presents AppQ, a framework that evaluates app quality using "inborn features" extracted directly from source code and UI structures. By modeling apps as attributed view graphs and employing a specialized graph-to-vector encoding method, AppQ achieves up to 85% accuracy in quality prediction, enabling effective recommendations before an app has even been downloaded.
Introduction
Most app store ranking and recommendation algorithms rely heavily on user-generated data: star ratings, review sentiment, and download totals. This creates a classic cold-start problem: high-quality new apps remain invisible because they lack the historical data needed to climb the rankings. Furthermore, user reviews can be unreliable, biased, or even manipulated. AppQ addresses this by measuring an app's quality based on its intrinsic characteristics—how it is built and how a user interacts with its interface—rather than just its popularity.
Methodology
The AppQ system analyzes an Android app's APK to extract features across two distinct levels: static and dynamic analysis.
1. App-level Features (Static Analysis)
The system decompiles the app to check for requested permissions, the presence of advertisement libraries, and the volume of multimedia files (images/videos). These features serve as indicators of developer intent and potential user annoyance (e.g., intrusive ads or excessive permissions).
2. View-level Features (Dynamic Analysis)
Using automated traversal tools, AppQ records the layout hierarchy of each screen (view) and the transitions between them. This captures the structural complexity and interactive flow of the application.
The Attributed View Graph and Encoding
The core innovation of AppQ is representing the app as a graph where nodes are views and edges are transitions (clicks) between them. To translate this complex graph into a format suitable for machine learning, the authors developed a custom Graph-to-Vector approach rather than using standard Graph Neural Networks (GNNs), which were found to be too noisy for this specific task.
As shown in the figure above, the encoding process involves three main steps:
- Key Vertex Selection: Graph centrality is used to identify the most important "hub" views in an app, filtering out noise from secondary screens like "About" or "Settings."
- Neighbor Assembly: The system captures the features of these hub views and their immediate neighbors to preserve local structural context.
- Ensemble Classification: These features are fed into an ensemble of three classifiers (Linear-SVM, Rbf-SVM, and Random Forest) to categorize the app as low, average, or high quality.
Key Results and Insights
- High Accuracy: AppQ achieved a peak accuracy of 85.0% in the Weather category, significantly outperforming baseline models.
- UI as a Proxy for Quality: The researchers found that high-quality apps tend to have more complex, layered layouts and a higher number of interactive widgets, suggesting that developer proficiency in UI design correlates strongly with overall quality.
- Efficiency: The entire analysis pipeline takes an average of 183 seconds per app, making it a feasible solution for app stores to integrate into their submission vetting processes.
- Superiority over GNNs: The custom graph-to-vector method outperformed state-of-the-art models like GCN and GAT by focusing on "main views" while ignoring unimportant structural noise.
Conclusion
AppQ demonstrates that an app's internal structure and UI design are powerful predictors of its eventual success and quality. By leveraging attributed view graphs, app stores can provide fairer recommendations for new developers and ensure that high-quality software reaches users faster, effectively solving the cold-start problem in the mobile ecosystem.
Summary updated