Title: Automated Assignment and Classification of Software Issues
Advisor: F. Başak Aydemir
Abstract: Software issues serve as units of work for development teams, facilitating communication and task allocation. Assigning issues to the most suitable team member and accurately categorizing them is a complex and time-consuming task. Misclassifications can lead to project delays, rework, and difficulties within the team. In this thesis, Büşra proposes a carefully curated set of linguistic features for shallow machine learning methods. The performance of these methods is compared with deep language models. The novelty lies in assigning issues to four roles (designer, developer, tester, and leader) instead of specific individuals or teams, enhancing the solution's generality. Additionally, the developers' level of experience is considered to reflect industrial practices. The thesis employs a classification approach to categorize issues into bug, new feature, improvement, and other, while further classifying bugs based on specific modifications required. Industrial data sets, consisting of 5324 annotated issues from a leading global television producer, are collected and evaluated. Büşra's research demonstrates that an ensemble classifier of shallow techniques achieves remarkable accuracy (0.92 for issue assignment and 0.90 for issue classification), statistically comparable to state-of-the-art deep language models. The contributions include the public sharing of five annotated industrial issue data sets and the validation of the efficacy of an ensemble classifier of shallow machine learning techniques.
