Förbättra Detektionen av Testinstabilitet med Logbaserade egenskaper : En maskininlärningsmetod

Flaky tests can take a lot of time off the software development process since developers can not trust that their test result is deterministic unless they rerun the test many times. This has led to an increased interest in finding or predicting flaky tests before running even a single test. Studies show promising results using machine learning models where flakiness is mostly predicted correctly and in some studies even analyzed to assist developers in fixing the flakiness in the test. However, utilizing information obtained from running tests can reveal additional information that can help in detecting flakiness. This thesis utilizes an embedding model based on TeleRoBERTa to embed important information of log files and the test code. These embedding together with metrics of the test code and test code smells, two datasets are created and compared. One with the log embedding and one without. Then comparing machine learning models trained on each dataset the impact of the log features is obtained. The result shows a decrease in prediction performance when the log information is added which is not as expected and must mean that the method is not suitable to extract important information from the log file. However, on the dataset without information from the log file, the performance is comparable with existing techniques to catch flakiness, showing that the method still works if applied correctly.