LinkRank Lifts One-to-Many Issue-Commit Traceability F1 From 48.39% to 74.54% by Ranking Instead of Pairwise Classification

arXiv 2607.23610·low signal

Most traceability recovery assumes one issue maps to one commit, but real issues are frequently resolved across several commits, so pairwise classification systematically produces incomplete links. LinkRank (arXiv 2607.23610) reframes the task as issue-centric ranking with an iterative pick–remove–renormalize loop: select the top-ranked commit, drop it from the candidate pool, renormalize the rest, repeat until a stopping criterion fires. On a new dataset from six open-source GitHub repositories it reaches 74.54% average F1 in the Known-K setting versus 48.39% for the strongest baseline, and in the harder Unknown-K setting still hits 68.84% F1 with an absolute stopping rule and 67.02% with a relative one, beating baselines under both.