What is entity resolution?
Entity resolution is the practice of recognising that several records refer to the same customer or supplier despite differing in form. Names are abbreviated and addresses are entered inconsistently, so matching cannot rely on equality. The output is a resolved identity that downstream retrieval and reporting can both depend on.
Also called: Record linkage
Why an answer about a customer is only as good as the match behind it
An answering layer asked about a customer's exposure will confidently total whatever records it matched, and will say nothing about the records it missed. In a bank that understates a position; in a hospital it separates a patient from half their history. The cost is a correct-looking answer with no indication of incompleteness, which is worse than a refusal because nobody investigates it. Resolution quality therefore sets a ceiling on every answer that aggregates across records.
- Aggregated answers are wrong by omission whenever a record fails to match.
- Incompleteness is invisible to the reader, since the answer looks whole.
- Over-matching is the opposite failure and discloses one party's data to another.
- Match decisions need a record, because a regulator can ask why two records were joined.
How entity resolution works
- Candidate records are standardised so that names and addresses take a common form.
- Blocking narrows comparison to records that could plausibly refer to one party.
- Pairs are scored on multiple attributes rather than on a single key.
- Scores above a confirmed threshold merge automatically, with the evidence retained.
- Scores in the uncertain band are referred to a person for adjudication.
- Merges remain reversible, since a split is sometimes required later.
How entity resolution is evaluated
Entity resolution is assessed on both directions of error. A missed match and an invented one carry very different consequences, so examine the following against an adjudicated sample:
- precision and recall of automatic merges against human adjudication
- share of pairs falling in the uncertain band
- rate of merges later reversed by a person
- match quality on records with a missing identifier
- coverage of the standardisation rules across each source system
Limitations, risks and common failure modes
- Common surnames and shared addresses produce merges of genuinely different people.
- A single strong identifier is trusted while the rest of the record contradicts it.
- Thresholds are tuned on a clean sample and applied to a corpus that is not clean.
- Merges are implemented destructively, so the original records cannot be recovered.
- Resolution is performed once at migration and never repeated as records arrive.