Probabilistic matching

Probabilistic matching estimates the likelihood that two records describe the same person by scoring how closely their attributes agree, then links pairs above a chosen threshold.

Part of the Learn hub and the customer data glossary.

What it means in practice

Where deterministic matching needs an exact shared key, probabilistic matching works from partial evidence. It compares attributes such as name, date of birth, address and phone number and, in marketing contexts, device and behavioural signals. Each comparison adds or subtracts weight according to how informative the agreement is: sharing a rare surname counts for more than sharing a common one.

The threshold is a policy decision, not a technical constant. Raising it reduces false merges and increases missed matches. Tune it on a labelled sample of your own data and revisit it when data sources change.

Because the method is inference, store the score with the link and let the permitted use of a link depend on its confidence. A low-confidence link might be allowed to suppress a message, which errs on the side of not contacting someone, but not to personalise content or disclose account information.

Example (healthcare): an enterprise master patient index matches registrations from several systems on name, date of birth and address. Typographical variants can make exact matching miss true duplicates, so scoring is used, with a high threshold and a review queue, because a wrong merge could place one patient's data in another patient's record.

Related terms

• Deterministic matching: the exact-key alternative.
• Identity resolution: the wider process.
• Identity graph: where scored links are stored.
• Golden record: the profile built after matching.
• Explainability: why scores need recorded reasons.

Where it matters

• Identity resolution explained: choosing thresholds and review bands.
• Patient 360 versus the EHR: why patient matching carries higher stakes.