EXPERIMENTAL RESEARCH ON APPROACHES TO GENERATING TEST SELECTORS USING GNN IN THE PROCESS OF AUTOMATED TESTING OF WEB APPLICATIONS

The article discusses the problem of instability of test selectors in the process of automated testing of web applications. It raises the issue of selectors’ adaptability to changes in the DOM structure, which is critically important in the development of modern dynamic web interfaces. A comparative analysis of three approaches to selector generation is conducted: manual (via Chrome DevTools), semi-automated (using DevTools), and automated using graph neural networks (GNN).
The aim of the study was to determine which approach provides the best balance between accuracy, completeness, stability, and time spent on selector formation in real testing conditions. The work used an experimental approach that included testing on 25 web applications with different types of DOM structures: static, dynamic, and SPA. Standard quality metrics were used: accuracy, completeness, F1-score, and the number of errors after changes in the DOM and the average selector generation time were also evaluated.
The results show the advantage of the GNN-based method, particularly in the stability of selectors after interface updates. Although the manual approach demonstrates high accuracy, it is significantly inferior in terms of speed and completeness, and Chrome DevTools, although fast, is the least reliable in dynamic environments. Thus, the use of graph neural networks makes it possible to create more adaptive and reliable solutions for automated testing of web applications.
In addition, the study demonstrates the feasibility of further developing GNN-based approaches through the integration of self-supervised learning and Graph Attention Networks (GAT) mechanisms, which will enable even greater accuracy, scalability, and performance of automated tests in complex software systems. The proposed approach has the potential to be implemented in next-generation frameworks for UI testing, which will contribute to improving the quality of digital products on the market.
Also, the feasibility of using GNN for automated generation of selectors adapted to changes in the DOM has been experimentally proven, and it has been found that this approach reduces the number of false positives without compromising performance. The results obtained are of practical value for improving the reliability and efficiency of automated testing systems in rapidly changing interface environments.