Improving Mobile Application GUI Testability with Deep Learning-based Test Case Generation

The proliferation of mobile applications created the need to automate graphical user interface (GUI) testing, and one notable practice is deep learning-based test case generation. However, the testability impact of this practice has yet to be explored. Testability and automation are strongly correlated, where low testability reduces the benefits of automation, while ineffective automation will negatively affect testability. This work aimed to explore the effect of deep learning-based GUI test generation on the testability of mobile applications. First, we compared four deep learning algorithms to classify mobile GUI elements (Detectron2, EfficientDet, YOLOv5, and YOLOv8). Comparison results showed that YOLOv8 outperformed the other models in precision, recall, and AP50 scores. Afterward, we applied test case generation on two Android applications where metrics defined by ISO/IEC 25023:2016 provide standards to measure testability. Evaluation results showed improvements in both applications' testability, where the generated test cases increased the conformance of the required test coverage. We noted at least six times improvement in testability. This work concluded that deep learning-based GUI test case generation could improve the testability of mobile applications by creating dozens of applicable test cases.