Entity Resolution
Entity resolution is the process of working out which records refer to the same real-world thing — the same person, business, or device — when no shared identifier says so. It is how “J. Smith, 14 Oak St” and “John Smith, 14 Oak Street Apt 2” become one customer, and how forty accounts become one fraud ring.
| Also called | Record linkage, identity resolution, deduplication, data matching |
| Core problem | The same entity appears in many records with no common key |
| Deterministic matching | Exact rules on chosen fields — fast, precise, brittle |
| Probabilistic matching | Weighted similarity across fields, producing a confidence score |
| Graph-based | Entities as nodes, shared attributes as edges — finds indirect connections |
| Two error types | False positives merge distinct people; false negatives split one person |
| Threshold effect | Tightening one error type always loosens the other |
| Fraud use | Linking accounts that individually look unrelated |
| Compliance use | Single customer view, sanctions matching, beneficial ownership |
How it works
Three families of technique, usually combined rather than chosen between.
Deterministic matching applies exact rules: same national ID number, or same surname plus date of birth plus postal code. It is fast, easy to explain to an auditor, and brittle. A transposed digit, a married name, or an address recorded two ways defeats it silently.
Probabilistic matching scores similarity across many fields and weights them by how much each contributes. A shared surname is weak evidence; a shared surname, date of birth and phone number is strong. The output is a confidence score, and where the threshold sits on that score is the whole decision.
Graph-based resolution treats entities as nodes and shared attributes as edges, which surfaces indirect connections the other two miss. Two accounts sharing nothing directly may both connect to a third through a device and a phone number — a relationship invisible to pairwise comparison and obvious in a graph. This is what finds organized fraud.
The error trade-off is unavoidable and worth stating plainly. False positives merge two real people into one record, which produces wrong decisions about both and is very hard to unpick afterwards. False negatives leave one person fragmented across several records, so the institution never sees the whole relationship. Tightening either loosens the other, and the right setting depends on which error costs more in that context.
Why it matters for identity verification
Entity resolution is what turns individually unremarkable applications into a visible pattern, and its quality depends entirely on what it has to match on.
A fraud ring submitting forty applications with forty distinct identities looks like forty unrelated customers under per-application scoring. Resolution is what connects them — and what it can connect them on determines whether it works. Device fingerprints and IP addresses are useful and cheaply varied by a competent operator using clean devices and residential proxies. Self-reported names and addresses are noisy and easily perturbed.
Verified identity attributes are the hard ones to vary. The same authenticated document appearing behind several applications, or the same face across identities that share nothing else, is a link a fraudster cannot cheaply break — because breaking it means acquiring another genuine document and another cooperating person. Linking verified identity attributes across accounts is what makes resolution robust rather than merely clever, and Microblink’s synthetic and stolen identity detection is built on that linkage.
The same machinery serves compliance. A single customer view, accurate sanctions matching, and beneficial ownership mapping are all entity resolution problems — and all of them fail in the same direction when the underlying identity data was never verified.
Deterministic vs probabilistic matching
| Deterministic | Probabilistic | |
|---|---|---|
| Basis | Exact rules on chosen fields | Weighted similarity across many fields |
| Output | Match or no match | A confidence score |
| Handles typos and variants | No | Yes |
| Explainability | High — the rule is the explanation | Lower; requires score interpretation |
| Tuning | Edit the rules | Move the threshold, retrain the weights |
| Typical failure | Misses real matches | Merges records it should not |
| Best suited to | Strong shared identifiers | Messy, incomplete, human-entered data |
Most production systems run deterministic rules first to catch the certain matches cheaply, then pass the remainder to probabilistic scoring. Graph analysis sits on top of both.
What it can’t do
It cannot invent information that is not there. Two records sharing only a common name cannot be resolved confidently, and no algorithm changes that. Resolution quality is bounded by the attributes available, which is why what gets captured at onboarding matters more than the matching technique.
It cannot avoid both error types. Every threshold trades false merges against missed links. A system claiming to eliminate both has simply chosen a setting and not told you which.
False merges are expensive to undo. Once two people are treated as one, decisions accumulate against a combined record — credit, risk scores, communications. Unpicking that afterwards is substantially harder than preventing it, which argues for caution on the merge side in consumer contexts.
It does not establish that any identity is real. Resolution determines which records refer to the same entity. Whether that entity is a real person is a different question, and a well-constructed synthetic identity resolves cleanly to itself.
Frequently asked questions
What is the difference between entity resolution and identity resolution?
Largely usage. Entity resolution is the general term and covers people, businesses, devices and products. Identity resolution usually refers specifically to people, and in marketing contexts to stitching a customer’s activity across channels.
What is the difference between deterministic and probabilistic matching?
Deterministic applies exact rules and returns match or no match — fast and explainable, but defeated by typos and variants. Probabilistic scores similarity across fields and returns a confidence value, handling messy data at the cost of a threshold decision.
How does entity resolution detect fraud rings?
By connecting accounts on shared attributes, including indirect connections found through graph analysis. Individually the applications look unrelated; in aggregate they share devices, addresses, or verified identity attributes that reveal coordination.
What makes entity resolution fail?
Thin or unverified attributes. Two records sharing only a common name cannot be resolved confidently, and self-reported data is easily perturbed by a fraudster. Verified identity attributes are far harder to vary, which is what makes the linkage hold.
Related reading
- Fraud ring — what entity resolution is usually looking for
- Synthetic identity fraud — the identity that resolves cleanly and is not real
- Identity document verification — where the hard-to-vary attributes come from
- Fraud detection — the layer resolution feeds
