Recommending Issue Reports to Developers Using Machine Learning

Abel Mesfin Cherinet
The development of a software system is often done through an iterative process and different change requests arise when bugs and defects are detected or new features need to be added. These requirements are recorded as issue reports and put in the backlog of the software project for developers to work on. The assignment of these issue reports to developers is done in different ways. One common approach is self-assignment, where the developers themselves pick the issue reports they are interested in and assign themselves. Practising self-assignment in large projects can be challenging for developers because the backlog of large projects become loaded with many issue reports, which makes it hard for developers to filter out the issue reports in line with their interest. To tackle this problem, a machine learning-based recommender system is proposed in this thesis. This recommender system can learn from the history of the issue reports that each developer worked on previously and recommend new issue reports suited to each developer. To implement this recommender system, issue reports were collected from 6 different opensource projects and different machine learning techniques were applied and compared in order to determine the most suitable one. For evaluating the performance of the recommender system, the Precision, Recall, F1-score and Mean Average Precision metrics were used. The results show that, from a backlog of 100 issue reports, by recommending the top 10 issue reports to each developer a recall ranging from 52.9% up to 96% can be achieved, which is 6 up to 9.5 times better than picking 10 issue reports randomly. Comparable improvements were also achieved in the other metrics.
Graduation Thesis language
Graduation Thesis type
Master - Software Engineering
Ezequiel Scott
Defence year