Skip to content

Identity Resolution Model

Inference model — instantiates Equivalence Class Consolidation

An inference model that weighs evidence across attributes to decide, with a confidence score, whether two records or names refer to the same real-world entity.

An Identity Resolution Model is the matching brain of a consolidation: given two records that might refer to the same person, company, product, or account, it decides whether they do — and how sure it is. It is the mechanism that establishes the equivalence criterion itself, by weighing agreements and conflicts across attributes rather than demanding an exact key. Its defining trait is that it produces a graded answer: a match probability, not a yes/no, because real-world identity data is dirty, partial, and sometimes deliberately varied. Every other mechanism in the archetype acts on an equivalence judgment; this one makes it, and hands out a confidence with it so the actions can be routed by risk.

Example

A bank must know whether a new account holder — "Robert A. Smith, born 1971, London address" — is the same person as an "R. Smith" flagged on a sanctions list and as a "Bob Smith" already in its customer base. Exact matching fails immediately: the names, formats, and identifiers all differ slightly. The identity resolution model instead scores the evidence. A fuzzy name agreement and a matching date of birth push the probability up; a shared prior address and phone push it further; a conflicting middle initial and a different national ID pull it down. It weighs each field by how discriminating it is — agreeing on a rare surname counts far more than agreeing on a common one.[1] The customer-base "Bob Smith" scores 0.93 and is auto-linked; the sanctions "R. Smith" scores 0.58 and is routed to a compliance officer for a manual decision rather than being frozen automatically. The model did not collapse the records or file a report — it produced the graded judgment that everything downstream depends on.

How it works

What sets this mechanism apart is that it reasons from weighted evidence and emits calibrated confidence:

  • It scores each candidate pair by how much each attribute discriminates identity — a match on a high-cardinality field carries more weight than a match on a common one, and conflicts carry negative weight.
  • It combines those weights into a match score, then applies thresholds that split pairs into confident match, confident non-match, and an in-between review band rather than forcing every pair to a verdict.
  • Because comparing every pair is infeasible, it blocks the population into plausible candidate sets first, then scores only within them.

Tuning parameters

  • Match / non-match thresholds — where the score cutoffs sit. Raising the match bar trades missed links (false non-matches) for fewer wrong links (false matches); the two errors have very different costs.
  • Attribute weighting — how much each field contributes and how conflicts are penalized. Mis-weighting a weak or correlated signal quietly inflates or deflates confidence.
  • Review-band width — how large the "send to a human" zone is. A wide band is safer but expensive; a narrow one automates more at the cost of unreviewed errors.
  • Blocking strategy — how aggressively candidate pairs are pruned before scoring. Tight blocking is fast but drops true matches that disagree on the blocking key.
  • Deterministic overrides — hard rules (a verified government ID match, or a known-distinct flag) that bypass the probabilistic score in either direction.

When it helps, and when it misleads

Its strength is resolving identity across data where exact keys are hopeless — typos, aliases, partial records, format drift, deliberate variation — and doing so with a confidence that lets high-risk links get human eyes while routine ones flow through.

Its gravest error is the false match: linking two genuinely different entities, which in this domain can freeze an innocent customer or, in a health or legal system, attach one person's record to another. Its mirror is the missed link. Confidence is only as trustworthy as the evidence's independence and quality: correlated attributes (a shared household address and phone) can be counted as if independent and inflate certainty, and adversaries deliberately vary the very fields the model leans on. Base rates compound the danger — when almost every candidate pair is a non-match, even an accurate model produces a flood of false positives in absolute terms. The misuse to guard against is tuning thresholds to hit a target match rate, or running the model to justify a link someone already wanted. The discipline is to calibrate thresholds against labeled ground truth, keep a genuine review band, monitor precision and recall against base rates, and treat every score as evidence to be adjudicated rather than a verdict.

How it implements the components

  • equivalence_criterion — the weighted-evidence rule and its thresholds are the criterion for when two records count as the same entity; the model defines it rather than applying a given one.
  • relevant_property_set — it selects and weights exactly the attributes that discriminate identity, separating meaningful sameness from surface resemblance.
  • confidence_level — its output is an explicit match probability, and routing by that confidence is intrinsic to how it works.

It does not sweep a whole population and collapse the matches into one — that is the Deduplication Workflow that consumes its scores — nor store the resulting variant map (Alias Resolution Table), nor build the governed surviving record (Master Record Consolidation).

  • Instantiates: Equivalence Class Consolidation — the model supplies the graded equivalence judgment the rest of the machinery acts on.
  • Sibling mechanisms: Deduplication Workflow · Master Record Consolidation · Alias Resolution Table · Canonicalization Pipeline · Crosswalk Table · Equivalence Test Suite · Policy Equivalence Rule · Synonym Merge Review · Taxonomy Merge Workshop · Unit Normalization Table

Notes

The model outputs evidence, not authority. Its scores are only actionable once a downstream mechanism decides what a given confidence licenses — auto-link, hold for review, or reject — and that mapping of confidence to action is where domain risk (a false AML freeze versus a missed sanctioned party) actually lives. Keeping the scoring separate from the action lets the model be recalibrated without re-arguing every policy threshold.

References

[1] The Fellegi–Sunter model of probabilistic record linkage formalizes exactly this: it weights agreement and disagreement on each field by its discriminating power and sets two thresholds to yield match, non-match, and a clerical-review region — the standard framework behind evidence-weighted identity resolution. ↩