Search papers, labs, and topics across Lattice.
This paper details a novel methodology for disambiguating author identities in the World of Code collection, achieving a significant increase in scale by resolving approximately 107 million author strings across 6 billion commits. The authors tackle the challenge of over-merging unrelated identities into large clusters, employing a combination of node-level gates and a per-edge classifier trained on extensive labeled data to enhance both recall and precision. The resulting system demonstrates a remarkable improvement in accuracy, reducing the size of the largest cluster from over 170,000 to under 7,000 while increasing gold recall from 0.44 to 0.70, outperforming previous methods in the field.
Overcoming the challenge of over-merging, this methodology achieves a 99% AUC in identity resolution while drastically reducing mega-cluster sizes in the World of Code dataset.
We describe the methodology used to alias the free-text author/committer identities of the entire World of Code (WoC) collection (version V2604, ~107M distinct author strings over ~6B commits) into canonical persons, extending the fingerprint-based anti-aliasing of ALFAA and the 38M-identity resolution of Fry et al. by an order of magnitude. At this scale the central problem is over-merge, not missed merges: a few bridge identities (bots, role accounts, placeholder emails, multi-author commit fields) transitively weld unrelated clusters through the global union step into million-member"mega-clusters."We report the full experimental record (more than twenty experiments, including unsuccessful ones) behind the deployed design. Node-level gates (information score, project spread, degree) preserve recall but cannot dissolve the mega-cluster; per-value blocklists of high-quality-but-shared attributes are recall-safe but cannot break a redundant mesh; the working composition is a betweenness cut over the exact union graph plus a per-edge classifier trained on 2.6M labels mined from GitHub no-reply identifiers. That classifier, filtering dormant cross-project shingle groups and joined by GitHub's own account assertions, then recovers the recall the precision work had foregone. Against human-adjudicated pairs the per-edge model transfers at AUC 0.99; end to end the largest cluster falls from 170,431 (and a predecessor's 3.0M) to under 7,000 w hile gold recall rises from 0.44 to 0.70 at increasing precision, and on an independent 21M-alias GitHub ground truth the final map outscores its predecessors and the published state of the art among global, privacy-preserving resolvers. The record doubles as a catalog of scale lessons: structural cuts do not transfer to edge sets they never saw, and recall-only and precision-only benchmarks invert verdicts unless read together.
