Test Data Glossary: Definitions for Entity Resolution, Fuzzy Matching, Golden Record, and Synthetic Data
Entity resolution, fuzzy matching, golden record, and synthetic data each carry a different meaning in a test-data context. This glossary defines all four, marks the boundary of each term, and links to the procedural pages where the work happens. Definitions cover the testing context only; this page does not cover how to implement matching algorithms or how to generate a dataset.
This page defines. The linked pages show how. Jump to fuzzy matching, entity resolution, synthetic data, or golden record.
Fuzzy matching
Fuzzy matching is the practice of identifying records that refer to the same entity despite differences in how the name or other fields are written. Common techniques include edit-distance algorithms (Levenshtein, Damerau-Levenshtein), phonetic algorithms (Soundex, Metaphone, Double Metaphone), token-set comparison (Jaccard similarity on word-level tokens), and blocking (grouping candidates by a shared prefix or hash before comparison). A fuzzy match returns a score, not a binary result, and the caller sets a threshold. The term is used interchangeably with approximate string matching in the data-quality literature.
For a labeled dataset to test your own fuzzy-matching logic, see the guide with downloadable sample and answer key: Fuzzy matching test data with answer key.
Entity resolution
Entity resolution is the task of determining which records in one or more datasets refer to the same real-world entity. It is also called record linkage, data matching, data deduplication, or identity resolution depending on the industry. The core pipeline is: block (reduce the pair space), score (run a similarity or ML model on each candidate pair), and cluster (assign records to entity groups). The result is a set of entity clusters, each with a canonical ID. A ground truth set is a collection of record pairs annotated as match or no match by a human reviewer; it is the reference against which precision, recall, and F1 are measured.
For generating test data to exercise an ER pipeline end to end, see: Entity resolution test data, practical guide.
Synthetic data
Synthetic data is data generated algorithmically to mimic the statistical or structural properties of real data, without containing real records. In a test-data context the goal is coverage, not fidelity: you need records that exercise every code path, edge case, and failure mode your system must handle.
For generating synthetic test data in practice, see: Synthetic test data guide.
Truth set
Updated 2026-08-06
A truth set is a collection of record pairs whose match or no-match label is fixed in advance by a known answer key, independent of the match rules being tested. The live entity-resolution entry above describes the human-annotated case; a truth set also covers the synthetic case, where the generation process itself records which rows it created as duplicates of one another, so the answer key exists before any reviewer looks at a pair. Each pair carries one explicit label: the two records are the same entity, or they are not. Without a truth set a match rule's score is an assumption. With one, precision, recall, and F1 become numbers you can compute and defend.
For how to define the truth keys a labeled dataset needs and build one for your own domain, see: Fuzzy duplicates and truth keys, a practical how-to.
Survivorship
Updated 2026-08-06
A survivorship rule is the logic an MDM system applies to pick which field value wins when multiple matched source records disagree: for example, most-recently-updated or highest-confidence-source. When match/merge logic groups several source records into one entity, those records often carry conflicting values for the same field. One system might have an old phone number; another might have a new one. A survivorship rule resolves the conflict deterministically, so the same inputs always produce the same golden record output. The rule is a decision a data governance team makes and documents, and it is not automatically correct just because it is applied consistently, which is exactly the kind of logic a labeled truth set can help validate.
For why survivorship and match rule choices need to be proven, not assumed, see: Why an unvalidated match rule is a guess, not a decision.
Golden record
A golden record is the authoritative, deduplicated representation of an entity after a match/merge process. When multiple source records are matched to the same entity, the MDM or CRM system produces one output record by applying survivorship rules (for example: take the most recently updated value for each field, or prefer the higher-confidence source). In testing, the golden record is the survivor of a merge operation; the test corpus must include both the survivor and the records that were merged into it so the match/merge logic can be validated.
For MDM match/merge and golden record testing methodology, see: MDM test data guide.
Frequently asked questions
What is entity resolution?
Entity resolution is the task of determining which records in one or more datasets refer to the same real-world entity. It is also called record linkage, data matching, or identity resolution. The standard pipeline is: block, score, and cluster. It differs from deduplication in scope: deduplication typically operates within one dataset; entity resolution typically operates across multiple sources.
What is the difference between fuzzy matching and entity resolution?
Fuzzy matching is a technique: it compares two strings and returns a similarity score. Entity resolution is a pipeline that uses fuzzy matching (among other techniques) as its scoring step. A high fuzzy-match score is evidence for a match but does not itself constitute a resolved entity group.
What is a golden record in master data management?
A golden record is the authoritative, deduplicated record for an entity after a match/merge process. When multiple source records are matched to the same entity, the MDM system produces one output record by applying survivorship rules (for example: take the most recently updated value for each field, or prefer the higher-confidence source).
Is synthetic test data the same as fake data?
The terms are used interchangeably in informal usage, but have a distinction in formal contexts. Fake data typically means data invented without reference to a target distribution. Synthetic data typically means data generated to mimic a real dataset's statistical properties. For software testing, the important property is coverage, not fidelity.
Why does fuzzy matching sometimes return incorrect results?
Naive string-distance algorithms fail on transliteration, hyphenated surnames, full name changes, OCR artifacts, organization DBA names, legal suffix variation, merger prefixes, subsidiary suffixes, and hard negatives (distinct entities that share an identical name). Each failure mode needs its own technique class. For a labeled corpus and answer key, see the fuzzy-matching section above.
What is a truth set?
A truth set is a collection of record pairs whose match or no-match label is fixed in advance by a known answer key, independent of the rules being tested. That key can come from a reviewer, or, for synthetic data, from the generation process that created the duplicates. It is what precision, recall, and F1 are measured against. See the truth-key how-to for building one.
What is a survivorship rule in MDM?
A survivorship rule is the logic an MDM system applies to pick which field value wins when multiple matched source records disagree over the same field, for example, most-recently-updated or highest-confidence-source. It is a decision a governance team makes and documents, not something that becomes correct merely by being applied consistently, and it can be validated against a labeled truth set the same way any other match logic can.
What changed in the golden record definition?
The golden record entry itself is unchanged. A delta note was added pointing to the new Survivorship entry, which now carries the field-level conflict-resolution mechanism (for example, most-recently-updated or highest-confidence-source) that the golden record entry references. See the Survivorship entry above for that logic in full.