Research / Paper

Deduplication and Probabilistic Record Linkage

Comparative study of Transformer-based, Gradient Boosted, and Fellegi-Sunter models for identity resolution.

Built during a summer research fellowship, this project combines record linkage, evaluation design, and applied identity reasoning.

Result

The composite scoring protocol — combining the three approaches rather than choosing between them — outperformed each individual model as well as common industry methods for the same task.

Why it matters

Record linkage decides whether two records describe the same person. Getting it wrong in either direction has real consequences: duplicates corrupt downstream analysis, and false merges collapse distinct people into one identity.