Scaling Author Identity Disambiguation to the World of Code: A Methodology

2026-07-09 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The World of Code (WoC) V2604 project developed a methodology to scale author identity disambiguation for ~107M distinct author strings across ~6B commits. This approach addresses the critical "over-merge" problem, where bridge identities (bots, role accounts) create million-member "mega-clusters." The solution involved an experimental record of twenty approaches, culminating in a design that uses betweenness-centrality cuts on the union graph and a per-edge classifier. This classifier was trained on 2.6M labels mined from GitHub no-reply IDs. The deployed map reduced the largest cluster from 170,431 (and a predecessor's 3.0M) to under 7,000, while increasing gold recall from 0.44 to 0.70 at higher precision. It also outperforms prior maps and published state-of-the-art global resolvers.

Key takeaway

For Research Scientists and ML Engineers building large-scale identity resolution systems, you should prioritize detecting and dissolving "over-merge" early. Implement structural cuts like betweenness centrality and per-edge classifiers using free labels from sources like GitHub no-reply IDs. This approach improves precision and recall, preventing fabricated super-developers and corrupted analyses in massive datasets.

Key insights

Scaling author identity disambiguation requires structural cuts and per-edge classification to combat over-merge in large codebases.

Principles

• Structural certificates do not transfer to unseen edge sets.
• Gradient boosting can shortcut-learn label-construction artifacts.
• Identity resolvers must be graded on both recall and precision.

Method

The method combines a two-phase pipeline with betweenness centrality cuts on the union graph and a per-edge logistic classifier trained on GitHub no-reply IDs, followed by shingle expansion.

In practice

• Use GitHub no-reply IDs to generate millions of free identity labels.
• Apply betweenness centrality to identify and cut load-bearing nodes.
• Blank bad high-quality attributes like placeholder emails at the source.

Topics

• Author Identity Disambiguation
• World of Code
• Graph Algorithms
• Machine Learning
• Code Repository Mining
• Over-Merge Detection

Code references

• bockthom/gitauthority

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• A Global Author-Identity Map for the World of Code:62.7M Developer Identities from 106.8M Author Strings over 5.87B Commits — Accurate global author identity resolution requires prioritizing precision to prevent massive over-merges.
• Intro to NLP with spaCy (3): Detecting programming languages — Manual labeling and iterative error analysis are crucial for refining rule-based NLP systems and uncovering critical domain context.
• When Symbol Names Should Not Matter: A Logistic Theory of Fresh-Symbol Classification — Fresh-symbol generalization in transformers depends on collision graph geometry, not just token diversity.
• Codexa at SemEval-2026 Task 13: Loss Engineering and Diverse Ensemble Strategies for Multi-Class Code Authorship Attribution — Loss engineering and diverse model ensembles significantly improve multi-class code authorship attribution, outperforming domain adaptation.
• Project-wise Comparison of Software Birthmarks Using Weighted Partial Similarity — Project-wise software birthmark comparison requires weighting module similarities and focusing on top matches to detect partial reuse and mitigate incidental similarity.
• Redact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification — Problem formulation, separating PII candidate proposal from contextual Redact/Keep review, outperforms model scale for local educational de-identification.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.