Sending resolved customer data from the warehouse to the tools that act on it, such as ad platforms, email and CRM, often through reverse ETL. In a composable CDP it comes after identity resolution, so audiences go out with one profile per customer instead of duplicates.
Active learning
A training approach in which the model picks the record pairs it is least sure about and asks a person to label them, so a small number of labels teaches it the most. In entity resolution it cuts labelling effort compared with labelling random pairs. See thresholds and active learning.
Agentic entity resolution
Entity resolution in which AI agents take part in the work, for example proposing match rules, triaging uncertain pairs or explaining clusters, while the matching itself stays governed and auditable. The term is also used for resolution that serves agents, giving them one resolved entity to act on. See entity resolution for AI.
Agentic MDM
Master data management in which AI agents carry out stewardship work, such as reviewing match suggestions, filling missing attributes and flagging anomalies, with people approving the changes. It depends on resolved entities with persistent identifiers, so agents act on the right record. See master data management.
B
Batch entity resolution
Resolving a whole dataset in one run, usually on a schedule. It suits the first resolution of historical data and periodic mastering; incremental matching then keeps the result current.
Beneficial owner
The person who ultimately owns or controls a company. Entity resolution links owner records across registries, so that stakes held under different spellings add up to one person. Sanctions rules such as the US OFAC 50 percent rule depend on that aggregation.
Blocking
Narrowing which record pairs are compared by grouping records on cheap keys, such as a phonetic surname plus a postcode, and comparing only within each group. Without it, comparisons grow with the square of the data. Blocking that is too loose is slow, and blocking that is too strict silently drops true matches, because a pair that is never compared can never be matched. See blocking at scale.
Blocking key
The value used to assign a record to a group during blocking, for example the first three letters of a surname combined with the year of birth. Most systems use several blocking keys and combine the candidate pairs they produce.
C
Candidate pair
A pair of records selected by blocking for detailed comparison. Only candidate pairs are scored, so any true match that blocking fails to pair up is lost.
Cartesian join
Joining every row of one table to every row of another. Comparing a table with itself this way produces n(n−1)/2 distinct pairs, so a million records give about 500 billion comparisons. Matching without blocking is effectively a Cartesian join, which is why blocking exists.
Chaining
A clustering error in which indirect links merge records that do not belong together: if A matches B and B matches C, a simple approach merges A, B and C even when A and C are clearly different. One wrong link can join two large clusters. Cluster-level scoring and constraints reduce the risk. Also called transitive closure.
Clerical review
Human review of pairs the system is unsure about. The term comes from the Fellegi–Sunter tradition, where scores between two thresholds are sent to people. Today it is part of data stewardship.
Cluster
The set of records judged to describe one real-world entity. A resolution run turns pairwise match decisions into clusters, and each cluster usually gets an identifier.
Cluster approval
A stewardship step in which a person reviews a cluster and approves or rejects it before it is committed. The decisions should be kept, so later runs do not override them.
Collective entity resolution
Resolving entities using relationships as well as attributes, for example linking two company records because they share directors and an address. Also called graph-based entity resolution.
Composable CDP
A customer data platform assembled from components on your own data warehouse or lakehouse, with storage, identity resolution, modelling and activation as separate layers, instead of a single bundled product. Identity resolution is its foundation. See composable CDPs and identity resolution.
Conditional independence assumption
In the Fellegi–Sunter model, the assumption that whether one field agrees is independent of whether another does, once you know if the records match. Correlated fields such as street and postcode violate it, which makes scores overconfident. See Fellegi–Sunter limitations.
Connected components
In a graph of matched pairs, the groups of records joined directly or through other records. Taking connected components is the simplest way to turn pairwise matches into clusters, and the usual source of chaining errors.
Crosswalk
A cross-reference table that maps each system's local ID for an entity to a shared identifier. The clusters produced by entity resolution are often used to build and maintain it.
Customer 360
A single, unified view of each customer across all systems. It depends on entity resolution to know which records belong to the same customer. See why customer 360 matters.
Customer data platform (CDP)
Software that collects customer data from many sources into unified profiles for segmentation and activation. Identity resolution decides which records belong to the same profile: packaged CDPs do it inside the product, and composable CDPs do it in your own warehouse. See CDPs and identity resolution.
D
Data matching
An umbrella term for comparing and linking data that does not share a clean key. Entity resolution, record linkage, deduplication and fuzzy matching are all forms of data matching.
Data platform
The system where an organisation stores and processes its data, such as Databricks, Snowflake, Microsoft Fabric, Google BigQuery or AWS. Warehouse-native entity resolution runs inside it, so records are matched where they already live. See warehouse-native entity resolution.
Data profiling
Examining data before matching to see completeness, formats, value distributions and anomalies, such as placeholder emails or default dates of birth. It shows which fields are worth matching on and what needs standardising.
Data stewardship
The people and processes that review, correct and approve data, such as deciding uncertain matches and settling conflicts between sources.
Dedupe
Short for deduplication: finding and removing duplicate records. See deduplication.
Deciding that records match from exact or near-exact agreement on chosen fields, under explicit rules such as “same email and same date of birth”. It is fast and easy to explain, but brittle when values are mistyped, missing or inconsistent. See deterministic vs probabilistic matching.
E
Edit distance
The number of single-character insertions, deletions or substitutions needed to turn one string into another, also called Levenshtein distance. A small distance means the strings are similar.
Embedding
A list of numbers (a vector) that represents the meaning of a piece of text or a record, so that similar items sit close together. Modern matching and retrieval systems use embedding similarity alongside string similarity.
EMPI
Enterprise master patient index: a system that links a patient's records across the systems of a healthcare organisation. Patient matching is a hard entity resolution problem, and a 2018 Pew report cited match rates as low as about 80% within facilities.
Entity
Any distinct real-world thing a business cares about: a person, company, product, supplier, drug, household or location. Each entity is described by attributes such as name, address and identifiers.
Entity linking
A natural language processing task that connects a mention in text to an entry in a knowledge base. It is a different task from entity resolution, which links records in data sources to each other.
Entity matching
Deciding whether two descriptions refer to the same real-world entity. It is the pairwise decision at the centre of entity resolution.
Entity resolution
Identifying and linking records, in one data source or many, that describe the same real-world entity when there is no reliable shared key. The output is clusters of records, usually with a persistent identifier and often a golden record. See The What and Why of Entity Resolution.
Exact matching
Treating values as the same only when they are identical, usually after normalisation. It suits trustworthy identifiers such as tax IDs, but on names and addresses it misses typos and variants.
Explainability
Showing why records were matched: which fields agreed, how strongly, and which pairs connected a cluster. It lets data teams trust, audit and correct the results. See explainability.
F
F1 score
The harmonic mean of precision and recall, used as a single measure of matching quality. It can hide which kind of error dominates, so look at precision and recall separately too.
False merge
Two different entities wrongly judged to be the same. In practice it is often the more damaging error: a marketing email reaches the wrong person, or a data subject request exposes someone else's data. Also called a false positive.
Fellegi–Sunter model
The 1969 probabilistic framework for record linkage. For each field it estimates how likely agreement is for matches and for non-matches, turns the ratio into a weight, and sums the weights. Two thresholds split pairs into link, possible link and non-link. See Fellegi and Sunter (1969).
Fraud ring
A group of people or accounts acting together, often behind slightly different names, addresses, phone numbers or devices. Entity resolution, combined with relationship analysis, exposes rings by linking records that share these details. See fraud, AML and KYC.
Fuzzy matching
Comparing values that are similar but not identical, such as typos and abbreviations, using similarity functions instead of exact equality. It is one comparison technique inside entity resolution, not the whole of it.
G
Golden record
The consolidated, best available version of an entity, built from a cluster of matched records by applying survivorship rules. Good implementations keep lineage so every value can be traced to its source.
H
Householding
Grouping people who live together or share a relationship into a household entity, for marketing, lifetime value and donor analytics.
I
ID stitching
Another name for identity stitching.
Identity graph
A data structure that links identifiers (emails, phone numbers, device IDs, customer IDs) and records to individuals. Edges mean “same person” or “related to”.
Identity resolution
Entity resolution applied to people: customers, patients, citizens. In advertising and customer data platforms the term can also mean linking cookies, devices and other identifiers to one individual. See entity vs identity resolution.
Identity stitching
Linking records that share an identifier you trust, such as the same email, login or device ID. It is fast and easy to explain, but it misses records with no shared key and can over-merge when identifiers are shared or are placeholders such as test@test.com.
Incremental matching
Processing only new, changed or deleted records against existing clusters, so the resolved data stays current without rerunning everything. The hard part is merging and splitting clusters while keeping identifiers stable. See incremental identity resolution.
J
Jaccard similarity
The overlap between two sets divided by their union, often applied to the words or character n-grams of two strings. It copes well with reordered words, as in company names.
Jaro–Winkler similarity
A string similarity measure suited to short strings such as names. It counts matching characters and transpositions, and gives extra credit when strings share a prefix. Scores run from 0 to 1.
K
Know your customer (KYC)
The checks a financial institution performs to verify a customer's identity and assess risk. Entity resolution keeps one resolved profile for each customer across onboarding and reviews.
Knowledge graph
A graph of entities and the relationships between them. Its quality depends on entity resolution: unresolved duplicates become duplicate nodes. See entity-resolved knowledge graphs.
L
Labelled pairs
Record pairs a person has marked as match, non-match or unsure. They are the training data for a learned matching model, and a held-back set doubles as a test of quality.
Lineage
A record of which source records and rules produced each value in a golden record. It makes merges explainable and reversible.
LLM-based matching
Using a large language model to decide whether two records match by prompting it with both. It needs little or no training data but costs more per pair, so it is usually reserved for ambiguous pairs after blocking and cheaper scoring.
M
m-probability and u-probability
In the Fellegi–Sunter model, m is the probability that a field agrees when two records are a true match, and u is the probability that it agrees when they are not. Agreement adds log₂(m/u) to a pair's score.
Master data
The core business entities shared by many systems, such as customers, suppliers, products and locations.
Master data management (MDM)
The discipline and technology for creating and maintaining authoritative master data. It includes matching, golden records, survivorship, stewardship and governance. Entity resolution is one capability inside it. See evolution of MDM tools.
Match score
A number that expresses how likely two records, or a record and a cluster, are to describe the same entity. Thresholds on the score decide match, review or non-match.
Match type
In Zingg, the setting that tells the model how to compare a field, such as FUZZY for names, EXACT for codes, EMAIL, NUMERIC, or DONT_USE for fields carried through without matching.
Missed match
Two records of the same entity that the system fails to link, leaving a duplicate. Also called a false negative. Low recall means many missed matches.
Model drift
The gradual loss of matching quality as data changes, for example through new sources, formats or customer populations. Monitoring match rates and reviewing samples catches it early.
N
N-gram
A run of n consecutive characters or words, such as the character trigrams smi, mit and ith in smith. Comparing n-grams makes matching tolerant of typos, and they are a common blocking key.
Nickname dictionary
A mapping from names to their variants (for example Elizabeth to Liz, Beth and Betty) so that matching treats them as equivalent. Dictionaries differ by culture. The original value should not be overwritten, so nicknames are best handled inside matching. See What is in a name?
Normalisation
Converting values to a consistent form before comparison: case, punctuation, phone formats, abbreviations and address components. Keep the original values for display and audit.
P
Pairs completeness
The share of true matching pairs that survive blocking. It sets an upper limit on overall recall.
Persistent identifier
An ID for a resolved entity that stays the same across runs, even as clusters merge and split, so downstream systems can rely on it. See ZINGG_ID.
Phonetic algorithm
A function that codes words by how they sound, so that variants such as Smith and Smyth share a code. Soundex, Metaphone and Double Metaphone are common examples, and most were designed for English names.
Precision
The share of declared matches that are correct. Low precision means false merges.
Probabilistic matching
Deciding whether records match from the combined weight of evidence across fields, instead of exact rules. It copes with missing and noisy values. The Fellegi–Sunter model is the classic example, and machine-learned classifiers are a modern one.
R
Real-time entity resolution
Resolving a record against existing entities as it arrives, for example at sign-up, checkout or a screening check, so the application gets an entity ID back within the request. It relies on fast lookups against data that is already resolved.
Recall
The share of true matches that were found. Low recall means missed matches and duplicate records.
Record linkage
Matching records across datasets, a term from statistics and public health. Halbert Dunn used it in 1946 (Dunn, 1946). It is often used interchangeably with entity resolution.
Reduction ratio
The share of all possible record pairs that blocking eliminates. A high ratio means far fewer comparisons.
Reference data
Shared lookup and classification values, such as country codes, currencies and product categories.
Rule-based matching
Matching with hand-written rules that combine field comparisons, such as same surname, same date of birth and a similar address. Rules are transparent but multiply quickly and break on variants nobody anticipated. See deterministic matching.
S
Sanctions screening
Checking customers and counterparties against government watchlists. It needs fuzzy name matching that handles transliteration and name order, and it balances missed matches against false alerts.
Similarity metric
A function that scores how alike two values are, such as edit distance, Jaro–Winkler, Jaccard or embedding similarity. Different fields need different metrics. See similarity metrics.
Soundex
A phonetic algorithm that codes names by how they sound, so that similar-sounding names share a code (Smith and Smyth both become S530). It is used in blocking and comparison and was designed for English names.
Standardisation
Converting values into a consistent structure before matching, such as parsing an address into house number, street, city and postcode, expanding St to Street, or formatting phone numbers. It overlaps with normalisation, and the originals should be kept for display and audit.
Streaming entity resolution
Resolving records continuously as they flow through a stream of events, for example from Kafka, instead of in scheduled batches. It is real-time resolution applied to a feed.
Supervised learning
Training a model on labelled examples. In entity resolution the examples are record pairs marked match or non-match, and the model learns how much each field comparison should count.
Survivorship
The rules that decide which attribute value goes into the golden record when matched records disagree: most recent, most complete, most trusted source, most frequent, or a combination.
T
Threshold
A score cut-off that separates match, review and non-match. The right setting depends on the use case, because a false merge costs more in some settings than a missed match.
Transitive closure
Merging records through chains of links. See chaining.
U
Unmerging
Splitting a cluster again when new evidence shows records were wrongly joined. It needs lineage and stable identifiers to avoid breaking downstream systems.
Z
ZINGG_ID
The persistent identifier Zingg Enterprise assigns to each resolved entity. It survives new data, reruns and retraining. See ZINGG_ID.
