Understanding and Predicting Typed Links in Issue Tracking Systems

Abstract

This talk summarizes two recent papers: “Beyond Duplicates: Towards Understanding and Predicting Link Types in Issue Tracking Systems” accepted at MSR 2022, and “Automated Detection of Typed Links in Issue Trackers” accepted at RE 2022. In issue trackers like JIRA, stakeholders connect issues via links of certain types, such as Epic-, Block-, Duplicate-, or Relate-links. While previous research focused on Duplicate-links, we aim at understanding and predicting other link types. In the MSR paper, we studied issues linking in 15 public JIRA repositories. We evaluated the robustness of state-of-the-art duplicate detection approaches on our dataset with diversified link types. We found that current deep-learning approaches confuse duplicates and other links. Extending the training sets with other link types partly solves this problem. In the RE paper, we trained and evaluated various machine learning models to detect typed links. We found that a BERT model trained on titles and descriptions of linked issues outperforms other deep learning models, achieving an average macro F1-score of 0.64. We also studied what impacts the prediction performance and found that this depends on how repositories are used (e.g. linking quality) and by whom.