Skip to main content
Entity resolution (ER) is the process of finding records that refer to the same real-world entity, whether within one dataset (deduplication) or across two (record linkage). GoldenMatch breaks this into four stages.

​
Blocking

Comparing every pair of records is quadratic and infeasible at scale. Blocking reduces candidate pairs by grouping records that share a key (for example, the same ZIP or the same soundex of a surname) and only comparing within blocks. The blocking key is the single biggest performance lever. For N rows split into blocks of sizes n1, n2, ...: 
candidate_pairs = sum_i (ni choose 2)
A good key produces many small blocks. A poor key (for example, blocking on a low-cardinality city field) produces a few huge blocks and an explosion of pairs. See scale envelope for the guard rails. GoldenMatch ships 8+ blocking strategies: static, adaptive, sorted_neighborhood, multi_pass, ann, ann_pairs, canopy, and learned (data-driven predicate selection).

​
Scoring

Within each block, candidate pairs are scored. A matchkey is a named rule that combines one or more fields, each with a scorer and a weight, and a threshold above which a pair is a match. GoldenMatch provides 12+ scoring methods, including:
  • String similarity: exact, jaro_winkler, levenshtein, token_sort, soundex_match, dice, jaccard.
  • Name-aware: name_freq_weighted_jw (surname IDF-weighted), given_name_aliased_jw (alias-aware).
  • Embedding: embedding, record_embedding.
  • Probabilistic: Fellegi-Sunter EM-trained m/u probabilities with automatic threshold estimation.
  • LLM: GPT-4o-mini or Claude scores borderline pairs, with budget caps and graceful degradation.

​
Clustering

Pairwise matches are transitively closed into clusters: if A matches B and B matches C, then A, B, and C form one cluster. Clusters are labeled strong, weak, or split. Oversized clusters are automatically split on their weakest edge using a minimum spanning tree, which prevents a single shared value (a common info@ email, say) from collapsing unrelated records into one giant cluster.

​
Survivorship

A golden record is the canonical record synthesized from a cluster. GoldenMatch offers five merge strategies:
  • most_complete
  • majority_vote
  • source_priority
  • most_recent
  • first_non_null
Fields can be weighted by source quality (from GoldenCheck), and field-level provenance tracks which source row contributed each value.

​
Privacy-preserving linkage

When you cannot share raw data across organizations, PPRL matches records using Bloom-filter encodings instead of plaintext, reaching F1 0.924 on the FEBRL4 benchmark.

Auto-config

GoldenMatch can choose all of the above for you and iterate until it converges.