Entity Resolution at Scale: How Graphs Match Millions of Records Without Losing Precision

Summary

• Entity resolution at scale is primarily a precision problem. As record volumes grow, the risk of incorrectly merging separate entities (over-resolution) or failing to connect records that belong together (under-resolution) increases without the right matching framework.
• Two failure modes define the challenge: over-resolution blends unrelated entities and can suppress fraud alerts, while under-resolution leaves coordinated behavior hidden across fragmented accounts and identities.
• Graph-based entity resolution adds relationship context to conventional attribute matching: instead of asking only whether two records look alike, it evaluates whether their connected accounts, devices, transactions, addresses, and organizations support the same conclusion.
• At enterprise scale, graph provides relationship evidence that strengthens ambiguous matches, exposes structural contradictions that prevent false merges, and supports continuous identity updates as new data arrives.
• TigerGraph supports real-time entity resolution for fraud, AML/KYC, Customer 360, and healthcare use cases, combining distributed graph architecture with the relationship context needed to maintain precision across millions of records and billions of relationships.

A customer opens an account with one email address, contacts support using another, replaces a phone, moves to a new address, and later signs in from several devices. Meanwhile, different systems record variations of the same name and maintain their own customer identifiers.

Multiply that fragmentation across millions of customers, businesses, accounts, devices, and transactions, and entity resolution becomes more than a data-cleaning exercise.

It becomes a scale problem.

But processing more records is only part of the challenge. The harder problem is maintaining precision as the number of possible matches increases. Match too conservatively, and one real-world entity remains fragmented across multiple profiles. Match too aggressively, and unrelated entities are incorrectly merged.

Entity resolution at scale therefore requires more than faster fuzzy matching. It requires enough context to determine whether records that look similar actually belong together.

Graph-based entity resolution adds that context by evaluating the relationships surrounding each record alongside conventional identity attributes. Instead of asking only whether two records look alike, organizations can determine whether their connected behavior supports the same conclusion.

You’ll learn:

• Why entity resolution gets harder at enterprise scale, and why throughput alone is the wrong metric
• How over-resolution and under-resolution each cause distinct downstream problems for fraud, AML, and customer intelligence
• How graph-based entity resolution adds relationship context to conventional attribute matching
• What an enterprise entity resolution architecture needs to support continuous, explainable identity matching at production scale

Why Entity Resolution Gets Harder at Enterprise Scale

Entity resolution identifies records from different systems that refer to the same real-world person, company, device, product, or other entity.

At smaller volumes, organizations can compare attributes such as names, addresses, emails, phone numbers, dates of birth, and account identifiers. As datasets grow, the number of potential record matches increases dramatically. With one million records, for example, an exhaustive approach could involve hundreds of billions of possible record pairs. Comparing every record against every other record quickly becomes impractical at enterprise scale.

Entity resolution systems, therefore, use techniques such as blocking or candidate generation to reduce the search space. Instead of comparing every record with every other record, the system identifies a smaller group of plausible matches before applying more expensive comparison logic.

That introduces an important trade-off. If candidate rules are too broad, the system still performs an enormous amount of computation. If they are too restrictive, potentially valid matches may never reach the matching stage.

The problem becomes even harder when source data is incomplete or inconsistent. A customer may use a shortened name. An address may be outdated. Phone numbers may be recycled. Family members may share an address or device. Businesses may operate through subsidiaries with different legal names.

In fraud and financial crime, ambiguity can be intentional. Fraudsters routinely alter identifying information or distribute activity across multiple accounts to prevent systems from establishing that seemingly independent records are connected.

Scaling entity resolution, therefore, means controlling the volume of comparisons without throwing away the evidence needed to make accurate decisions.

Precision Matters More Than Simply Finding More Matches

A successful entity resolution program should not aim to merge as many records as possible. It should make the largest number of defensible matches.

Two failure modes make that distinction critical.

Over-resolution merges different entities

Over-resolution happens when records belonging to separate real-world entities are incorrectly combined.

Imagine two people with similar names who once lived at the same apartment complex. If an entity resolution model gives too much weight to name and address similarity, their records could be consolidated even though their accounts, devices, transactions, and histories are unrelated.

That error propagates downstream. Fraud scores may blend legitimate and suspicious behavior. Compliance history belonging to one party may be assigned to another. Customer profiles can contain contradictory activities. Models may calculate behavioral baselines using events that never belonged to the same person.

Merging unrelated identities can dilute risk signals enough to suppress fraud alerts, even when the downstream detection logic is functioning correctly.

Under-resolution leaves one entity fragmented

The opposite problem occurs when an organization fails to connect records that actually describe the same entity.

For fraud teams, that can make coordinated behavior appear to come from unrelated accounts. AML investigators may repeatedly review fragments of the same counterparty. Customer intelligence teams may see separate profiles for one customer across web, mobile, loyalty, and support channels.

The result is an incomplete understanding of the entity. Fragmented data creates costs that extend far beyond data management. 

When records remain disconnected across systems, the impact can also include:

• Fraud patterns that remain hidden across separate accounts and identities
• Incomplete customer profiles that limit service and personalization
• Duplicate marketing efforts directed at the same person through different records
• Compliance gaps caused by incomplete KYC information
• Supply chain risks that remain obscured by disconnected supplier and ownership data

This is why entity resolution at scale cannot be judged only by throughput. Match quality determines the quality of every fraud decision, risk score, customer profile, and analytical model built on top of it.

How Graph-Based Entity Resolution Changes the Matching Model

Conventional entity matching techniques, including deterministic and probabilistic matching, generally ask: How similar are these two records?

Graph-based entity resolution adds a second question: Do the relationships surrounding these records support the conclusion that they represent the same entity?

These approaches complement one another. Attribute-based matching identifies likely candidates, while graph analysis adds relationship context that can strengthen or challenge the match.

Start with strong deterministic evidence

Exact identifiers remain valuable when they are trustworthy. These might include:

• Verified customer identifiers
• Government-issued identifiers
• Verified email addresses
• Verified phone numbers
• Account or policy numbers

A strong exact match can quickly resolve many records without requiring complex analysis.

Add probabilistic attribute evidence

Not every record contains an authoritative identifier. Matching systems therefore evaluate attributes such as names, addresses, dates of birth, company names, email variations, and other descriptive fields.

Instead of requiring exact equality, probabilistic approaches calculate how strongly the available evidence supports a potential match.

But attribute similarity alone can be ambiguous. Two people can share a name. Several customers can live at one address. A business phone number can belong to many employees.

This is where graph adds another layer of evidence.

Evaluate relationship context

Suppose two customer records contain slightly different spellings of the same name. Attribute matching suggests that they might belong together.

The system can then evaluate whether those records are also associated with the same:

• Account
• Device
• Address
• Phone number
• Transaction history
• Organization
• Household
• Merchant relationship

Several consistent relationships can strengthen the match. Now consider two records with very similar names and addresses but entirely different account histories, devices, transaction patterns, and connected organizations. That surrounding structure provides evidence against merging them.

Preserve why a decision was made

A matching score alone may tell an investigator that two records were merged with 92% confidence. It does not necessarily explain why.

Graph-based entity resolution can retain the underlying evidence: which identifiers matched, which relationships supported the decision, and which connected records contributed to the result.

That matters when resolution decisions affect fraud alerts, audit, compliance investigations, or customer treatment.

How Graphs Preserve Precision as Entity Resolution Scales

Relationship context becomes increasingly valuable as datasets grow because high-volume environments naturally contain more ambiguous identifiers. Graph helps organizations manage that complexity in several ways.

Relationship evidence strengthens ambiguous matches

A similar name may provide weak evidence by itself. That same name combined with a shared verified phone number, account, device, and transaction pattern produces a much stronger case.

The matching process can evaluate multiple connected signals instead of depending disproportionately on one field.

Multiple signals reduce false merges

Graph can also prevent weak similarities from becoming incorrect matches. A shared address, for example, does not necessarily mean two customer records represent the same person. It may represent a family home, apartment building, corporate address, or assisted-living facility.

Connected context can reveal that the records otherwise belong to clearly separate activity patterns.

Structural contradictions expose over-resolution

Once records are merged, graph analysis can continue testing whether the resulting entity remains internally coherent. Separate device clusters, conflicting activity patterns, incompatible histories, or disconnected account groups may indicate that the system merged records too aggressively.

This is particularly important for fraud detection because an incorrectly consolidated entity can hide the very risk the organization is trying to detect.

Identity can evolve continuously

Entity resolution is not a one-time cleanup project. People move, replace devices, open new accounts, change names, and adopt new contact information. Businesses change ownership, reorganize, merge, and create subsidiaries.

New evidence can strengthen or weaken earlier matching decisions. An operational entity resolution system should therefore update continuously as data arrives rather than depending entirely on periodic batch reconciliation.

Parallel processing supports larger connected datasets

At enterprise scale, maintaining this context requires infrastructure capable of analyzing large numbers of entities and relationships without forcing matching logic to become artificially shallow.

TigerGraph uses distributed architecture and massively parallel processing to analyze highly connected enterprise datasets at scale. For entity resolution, that architecture supports analysis across large identity networks while retaining the relationship context needed for precise matching. Its entity-resolution solution is designed for real-time resolution across enterprise data rather than treating identity matching as an isolated batch process.

Entity Resolution at Scale Across Enterprise Use Cases

The importance of precision becomes particularly clear when resolved entities feed operational decisions.

Fraud and Synthetic Identity Detection

Fraudsters benefit when organizations cannot determine that apparently independent accounts are connected. A synthetic identity may combine legitimate and fabricated information, while fraud rings can reuse devices, addresses, phone numbers, beneficiaries, or payment instruments across multiple accounts.

Graph-based entity resolution connects these signals into a broader identity picture. Once records are correctly resolved, fraud detection systems can evaluate coordinated behavior across connected accounts rather than treating every profile independently.

TigerGraph’s fraud solution specifically focuses on analyzing relationships among accounts and entities to expose patterns that individual-record analysis can miss.

AML, KYC, and KYB

Financial institutions cannot evaluate customer or counterparty risk accurately until they know which records belong to which real-world entity. Entity resolution can help connect customers and businesses to accounts, addresses, directors, beneficial owners, counterparties, and other identifying evidence.

That resolved identity becomes the foundation for transaction monitoring, sanctions screening, beneficial-ownership analysis, and investigations.

If the underlying entity is fragmented, suspicious activity may remain distributed across apparently unrelated parties. If it is over-resolved, unrelated activity can contaminate a risk profile. For AML and KYC, identity precision is therefore part of risk precision.

Customer 360 and Household Resolution

The same challenge appears in customer intelligence. A consumer can appear separately in CRM, ecommerce, loyalty, mobile, service, and transaction systems. Consolidating those records is necessary, but a useful Customer 360 also needs to preserve household, device, product, journey, and behavioral relationships.

TigerGraph graph-powered Customer 360 is a continuously queryable network rather than a flat, periodically consolidated customer record. That allows identity resolution to become part of real-time personalization, retention, recommendations, and customer decisioning.

Healthcare and Patient Identity

Healthcare organizations face a similar problem when patient information is distributed across claims, encounters, prescriptions, providers, and administrative systems.

Fragmented identities can separate clinical and claims information that belongs together, while incorrect merges can associate care histories with the wrong individual.

Accurate entity resolution helps create a consistent identity foundation from which healthcare organizations can analyze patient journeys, provider relationships, fraud, and outcomes.

What an Enterprise Entity Resolution Architecture Needs

Entity resolution at scale requires more than a matching algorithm. Enterprises need an architecture that supports the full identity lifecycle.

Attribute and relationship matching: Traditional deterministic and probabilistic matching should work alongside connected context rather than being discarded.

Efficient candidate processing: The system must reduce unnecessary comparisons without creating filters so restrictive that genuine matches disappear before evaluation.

Continuous updates: New accounts, devices, addresses, transactions, and other evidence should update existing identities as they arrive.

Confidence and ambiguity management: Not every potential match should produce an automatic merge. High-confidence matches can be resolved automatically, clearly incompatible records can remain separate, and ambiguous cases can be routed for additional review.

Explainability and lineage: Investigators should be able to determine which records were connected, what evidence supported the resolution, and how the entity changed over time.

Production scalability: Matching must remain viable as record counts, relationships, incoming data, and concurrent analytical workloads grow.

These requirements shift entity resolution from a periodic data-quality exercise into persistent enterprise infrastructure.

From Entity Matching to Persistent Relationship Intelligence

Resolving records is valuable, but the larger opportunity comes from what the organization can do with the resulting identity.

Once several records are confidently associated with the same customer, company, device, provider, or counterparty, that entity becomes a reliable foundation for downstream analytics.

Fraud teams can see coordinated activity across accounts. Compliance teams can trace relationships around counterparties and beneficial owners. Customer teams can recognize people across channels and understand households and product relationships. AI systems can retrieve more complete connected context instead of reasoning over fragmented records.

That is the difference between simply deduplicating data and creating persistent relationship intelligence. The resolution process establishes who or what the entity is. The graph preserves how that entity connects to everything around it.

For enterprises operating across millions of records and billions of relationships, both capabilities matter.

Start with TigerGraph’s free tier or request a demo to bring entity resolution into production.

FAQs

What is entity resolution?

Entity resolution is the process of determining whether records from different systems refer to the same real-world person, organization, device, product, or other entity and linking or consolidating those records. At enterprise scale, this applies across millions or billions of records and relationships while maintaining the accuracy, performance, explainability, and operational reliability required for fraud detection, compliance, customer intelligence, and other identity-critical workloads.

Why is entity resolution difficult at scale?

As record volumes grow, the number of possible matches increases rapidly. Missing data, common identifiers, inconsistent attributes, and deliberately manipulated identity information also increase the risk of false merges and missed matches. Systems must narrow candidate comparisons without losing valid evidence, while managing continuous data updates and the ambiguity that comes from shared attributes across genuinely different real-world entities.

How do graphs improve entity resolution?

Graphs supplement attribute matching with relationship context. They allow a matching system to determine whether candidate records share consistent connections to accounts, devices, addresses, transactions, organizations, and other entities before deciding whether they belong together. This additional layer of evidence can strengthen ambiguous matches and prevent weak attribute similarities from producing incorrect merges.

What is the difference between entity resolution and identity resolution?

Entity resolution applies broadly to people, companies, devices, products, and other real-world entities. Identity resolution generally refers to the subset of entity resolution focused on person-, account-, or device-level identity across systems and channels.

Why is entity resolution important for fraud and AML?

Fraud and financial crime often span multiple accounts, identities, devices, counterparties, and organizations. Accurate entity resolution connects those fragments so fraud and AML systems can analyze the complete relationship network instead of evaluating isolated records. Without precise entity resolution, coordinated fraud activity may appear to come from unrelated parties, and suspicious counterparties may remain fragmented across compliance systems.