Final Thesis: Classification of Software Developer Contributions

Abstract: Software repositories contain large amounts of information about the evolution of a software project, much of which is documented through commits in version control systems such as Git. However, commit messages are frequently written in an inconsistent or incomplete way, making it difficult to analyse contributions systematically. This thesis investigates the automatic classification of software developer contributions based on GitHub commit data using deep learning methods. A prototype is developed that automatically constructs labelled training datasets from repositories that adopt the Conventional Commits specification, extracting the commit type label directly from the commit message and using the raw code diff as the input signal for classification. The same pipeline is applied to a complementary task of classifying GitHub issues into five categories based on their labels. Both datasets are used to fine-tune a DistilBERT model, which is evaluated on a held-out validation split using accuracy and macro-averaged F1 score. The commit classifier achieves 71.3% accuracy and a macro-F1 of 0.36 on a twelve- class taxonomy derived from the Conventional Commits specification. The issue classifier achieves 59.1% accuracy and a macro-F1 of 0.32 on a five-class taxonomy. The results show that classification is feasible for frequent and semantically distinct classes such as feat and fix, while minority classes such as perf and revert remain difficult to classify reliably due to insufficient training data and semantic overlap with adjacent classes. The findings demonstrate that transformer-based deep learning is a promising approach for commit classification, and that the Conventional Commits specification provides a scalable, annotation-free labelling strategy for dataset construction. At the same time, the pronounced gap between accuracy and macro-F1 highlights class imbalance as the primary challenge to be addressed in future work.

Keywords: None

PDF: Master Thesis

Reference: Ediz Kocak. Classification of Software Developer Contributions. Master Thesis. Friedrich-Alexander-Universität Erlangen-Nürnberg: 2026.