Glossary
Plain definitions for the terms used in identity resolution, audiences and agentic data work.
- Identity resolution
- Deciding whether different records describe the same real person, household or business, then linking them under one identity.
- Entity resolution
- The general form of identity resolution, applied to any kind of entity: people, businesses, products or places.
- Record linkage
- The technical task of finding records that refer to the same entity across or within datasets.
- Deterministic matching
- Linking records when specific fields match exactly or near exactly, such as an email or a phone number.
- Probabilistic matching
- Estimating the likelihood that two records match from the agreement of many fields. The Fellegi-Sunter model is the classic method.
- Semantic matching
- Comparing what records mean rather than how their text reads, using vector embeddings, so variants like Jo Jo and Johnny can converge.
- Contextual record linkage
- DarkMath’s term for linking records using context such as contact history, moves, vehicles and behavior, in addition to identifiers.
- Vector embedding
- A list of numbers that describes the meaning of a record. Records with similar meaning have embeddings that sit close together.
- Semantic Gravity
- DarkMath’s name for the way related signals cluster around one identity in vector space, pulling new data toward the profile it fits best.
- Golden Record
- The single, auditable profile for one resolved identity, with the sources and reasons behind every merge.
- Household
- People who share a residence and often share purchase decisions, built from resolved individuals.
- Persistent ID
- An identifier that stays stable for a person or household as their records change over time.
- Match rate
- The share of records linked to at least one other record. It can rise from bad guesses, so read it with precision.
- Precision
- Of the links a system made, the share that are correct.
- Recall
- Of the true links in the data, the share the system found.
- F1 score
- The harmonic mean of precision and recall. One number that rewards finding links without making false ones.
- False merge
- Two different people combined into one identity, such as a father and son at the same address.
- Missed link
- Two records of the same person left apart, so one customer looks like two.
- Cognitive Data Graph
- DarkMath’s connected model of resolved people, households, attributes, events and outcomes, each with its source, time, confidence and permitted use.
- Semantic attribute
- An inferred characteristic, such as life stage or media habits, derived from patterns in other data rather than read from one field.
- Seed audience
- The list of existing customers a lookalike model learns from.
- Semantic lookalike
- A person or household whose behavior and context resemble a seed audience in vector space.
- Synthetic data
- Realistic records generated for testing, demonstrations or training, clearly labeled and kept separate from real people.
- The Corruptor
- DarkMath’s controlled perturbation engine. It adds realistic errors to copies of records with known identities so matching methods can be scored against the truth.
- Clean room
- A secure environment where two parties analyze combined data without either seeing the other’s raw records.
- Anonymous ID
- A key that links records without carrying personal information, used to connect data between parties.
- Re-identification gate
- The control point where an anonymous result can be tied back to a person, allowed only with the data owner’s authorization.
- Contact coverage
- For a given audience, the share of records reachable by each channel, such as postal address, email or mobile ad ID.
- Hashed email
- An email address converted into a fixed code by a one-way function, used to match records without exchanging the address itself.
- Agentic workflow
- A sequence where software plans steps, uses tools and prepares actions toward a goal, with people approving what runs.
See the terms in action.
The DarkMatch demo shows a missed link and a false merge on five ordinary records.