Knowledge Graphs and Entity Resolution: Why Identity Needs Relationships, Not Just Similarity Scores

In Brief

Entity resolution and knowledge graphs solve related but different problems. Entity resolution determines whether records from different sources represent the same real-world entity. A knowledge graph describes how that resolved entity relates to other entities, classifications, hierarchies and business context.

This distinction matters because similarity is only evidence. Two records with similar names can represent different organizations, while records with very different names can represent the same legal entity. Relationships can provide additional evidence, but they should not automatically override weak identity resolution. A wrong merge can contaminate every relationship attached to the resulting entity.

The stronger architecture is therefore sequential: resolve identity, establish a persistent anchor, connect governed relationships, qualify the authority of those relationships and then serve the resulting context to applications and AI. Knowledge graphs become valuable when they enrich trusted identity—not when they are used to compensate for identity that was never resolved properly.

“International Business Machines Corporation” and “IBM” may refer to the same organization.

Two companies with almost identical names may be entirely different legal entities.

One customer may appear under a legal name in ERP, a commercial name in CRM and an abbreviated name in a service platform.

This is why enterprise identity cannot be reduced to string similarity.

Similarity is evidence, not identity.

Matching and entity resolution are not the same thing

The terms matching, deduplication and entity resolution are often used interchangeably.

For architecture purposes, it is useful to separate them.

Capability Question Output
Matching How similar are these records according to selected attributes and rules? Match evidence, comparison score or candidate pair.
Entity Resolution Do these records represent the same real-world entity? A resolved entity and the source records associated with it.
Knowledge Graph / Relationship Model How is that entity connected to other entities and business concepts? Governed relationships, hierarchies and contextual connections.

IBM MDM illustrates this separation clearly. Its matching algorithms compare source records and collect matching records into master-data entities. Relationship capabilities then connect records and entities through separately defined relationship types.

Matching produces evidence. Entity resolution establishes identity. Relationships provide context around that identity.

Why fuzzy matching remains useful

It is a mistake to describe fuzzy matching as an obsolete technology that knowledge graphs replace.

Attribute comparison remains fundamental to most entity-resolution systems.

Names, addresses, registration numbers, phone numbers and other attributes provide important evidence. Modern matching can combine deterministic rules, probabilistic techniques, machine learning and other similarity methods.

IBM's current MDM matching process, for example, includes standardization, bucketing and comparison and allows different algorithms to be configured for different entity types.

The limitation is not that attribute matching is useless. The limitation is that attribute similarity cannot by itself express the full business identity of an entity.

Similar names can describe different entities

Two organizations can share a brand, parent company or naming convention without being the same legal entity.

A highly similar name can therefore create a false positive.

The reverse also occurs.

An acquisition, legal-name change, abbreviation, transliteration or local naming convention can make records for the same organization look very different.

The problem becomes more difficult across countries and languages.

The closer entity resolution gets to legal and operational identity, the less safe it becomes to interpret similarity as equivalence.

Persistent identity is the anchor

Once records are resolved, the enterprise needs a stable way to refer to the entity.

Source identifiers may change when systems are replaced. A CRM customer number is not necessarily durable enterprise identity. Neither is a supplier number created inside one ERP instance.

A persistent enterprise identifier provides an anchor that applications and AI systems can reuse while source representations continue to change.

Source Records → Entity Resolution → Persistent Enterprise Identity

This is the identity foundation described in MDM #2, AI-Ready Master Data: Why Trusted Business Context Matters More Than Clean Data .

Relationships add context after identity is anchored

A resolved supplier entity becomes more useful when the enterprise can see its relationships.

Which parent company owns it? Which approved locations belong to it? Which products or materials does it supply? Which other entities sit inside the same corporate group?

A customer entity may connect to accounts, households, subsidiaries, service contracts or products.

Product entities may connect to components, categories, substitutes, suppliers and lifecycle states.

Entity resolution establishes the governed node. The knowledge graph explains the governed neighborhood around that node.

Relationship context can improve identity decisions

Although identity should anchor the graph, relationships can also provide evidence during entity resolution.

Suppose two supplier records have moderately similar names.

The case becomes stronger if both reference the same verified registration identifier, approved location or corporate parent.

Conversely, a seemingly strong name match becomes weaker if the records belong to incompatible legal hierarchies or countries and carry conflicting authoritative identifiers.

Relationship evidence therefore can improve adjudication.

Relationships can strengthen identity evidence. They should not silently redefine identity.

A graph can amplify a bad merge

Knowledge graphs are sometimes presented as though more connectivity automatically produces better understanding.

That is not always true.

If two distinct legal suppliers are incorrectly merged, the resulting entity may inherit contracts, products, locations and corporate relationships from both.

The graph then gives downstream applications a richer—but wrong—context.

AI systems can make the problem even more consequential because they may traverse and reason over those relationships automatically.

A knowledge graph does not repair weak entity resolution. It can amplify the consequences of weak entity resolution.

Govern relationships as carefully as attributes

Relationships should not automatically be treated as facts merely because they appear in a graph.

A parent-company relationship may come from an authoritative registry. Another may be inferred from external information. A customer-household relationship may be approved through an operational process. Another may be generated analytically.

The graph therefore needs governance state just as attributes do.

Relationship property Question
Type What kind of relationship exists between the entities?
Authority Is the relationship approved, observed, inferred, proposed or unresolved?
Provenance Which source, rule or decision established the relationship?
Effective Time When did the relationship become valid, and when did it end?
Evidence What information supports the relationship?

A graph edge is a data assertion. Important assertions need ownership, provenance and lifecycle.

IBM's current architecture shows the distinction clearly

IBM Master Data Management currently treats matching and relationship discovery as related but separate capabilities.

Matching algorithms analyze records and form master-data entities for deduplication and entity resolution.

Relationship discovery rules can then establish record-to-record, record-to-entity and entity-to-entity relationships according to defined conditions.

IBM also preserves metadata showing which rule created an automatically discovered relationship.

This is an important pattern: identity and relationships can reinforce one another without collapsing into the same control mechanism.

Current SAP architecture is moving toward the same connected-context model

SAP's 2026 Business Data Cloud architecture places native master-data management, governed semantics and knowledge graphs inside a common knowledge core.

SAP describes entity resolution as bringing records from SAP and non-SAP systems into trusted entity views, while knowledge-graph capabilities map relationships across data, entities and business processes for contextual understanding.

Following SAP's completed acquisition of Reltio, SAP Reltio is also positioned as a cloud-native MDM layer that combines entity resolution, relationships, hierarchies and interactions into multidomain entity views.

The architecture is evolving from “one golden record” toward “one governed entity surrounded by reusable business context.”

Knowledge graph is a logical capability first

Another common mistake is assuming that implementing relationship-aware MDM always requires introducing a dedicated graph database.

It may.

But the architecture decision should follow the use case.

Some MDM platforms persist entity relationships internally. Others expose graph-oriented views over existing entity models. An enterprise may also use a separate knowledge-graph platform when relationship traversal, semantic reasoning or cross-domain knowledge is broader than the operational MDM requirement.

SAP Business Data Cloud itself illustrates a multi-model direction, combining relational, graph, vector and other processing rather than requiring all business information to be moved into one graph store.

Knowledge graph is a logical capability first. Graph database is an implementation decision.

Do not assume that graph traversal equals reasoning

A graph makes relationships accessible.

It does not automatically determine what those relationships mean for every business decision.

An ownership path may matter to supplier risk but not to delivery routing. A household relationship may matter to marketing but be inappropriate for an individual credit decision. An inferred product association may support recommendation while remaining unsuitable for regulatory reporting.

Context becomes useful when the consuming process understands which relationships matter and which authority they carry.

Relationship discovery should preserve provenance

Automatically discovered relationships can be valuable, but the system should retain how they were created.

IBM's current relationship-discovery implementation, for example, records the rule identifier associated with an automatically created relationship and labels rule-generated relationships distinctly.

This is more than technical metadata.

It allows consumers and Stewards to distinguish a relationship created by an approved rule from one entered manually or derived through another method.

Relationship provenance allows the enterprise to ask not only “How are these entities connected?” but “Why do we believe they are connected?”

Human review remains important at the identity boundary

Matching systems will always encounter ambiguous cases.

IBM MDM, for example, supports a potential-match workflow in which comparison scores falling into an uncertain range become Data Steward tasks rather than automatic identity decisions.

Stewards can review side-by-side records, examine entity structure and choose to link, unlink or skip the case.

This is important because the cost of false positives and false negatives is rarely symmetrical.

Entity resolution should optimize for business risk, not simply maximize the number of records automatically linked.

A false merge can be more damaging than a duplicate

Duplicate records create fragmentation.

Incorrect merges create false identity.

The second can be more difficult to discover because the data appears cleaner after the merge.

Once the merged entity acquires relationships, survivorship decisions and downstream references, reversing the mistake can become significantly harder.

This connects directly to the controlled-recovery principles developed in MDM #4, Self-Healing Master Data: From Error Detection to Controlled Recovery .

A duplicate leaves one entity fragmented. A false merge can create one entity that never existed.

Measure entity resolution as a decision system

One accuracy percentage rarely describes entity-resolution quality adequately.

Useful evaluation should consider several dimensions.

Measure What it reveals
False-match pattern Where separate real-world entities are incorrectly combined.
Missed-match pattern Where multiple records for the same entity remain unresolved.
Review workload Whether uncertain cases create a sustainable Steward workload.
Decision reversals Where previous identity decisions are later found to be wrong.
Relationship integrity Whether relationships remain valid after entity merges, splits and lifecycle changes.

The objective is not zero duplicates at any cost. It is trustworthy enterprise identity at an acceptable level of business risk.

Do not use one matching policy for every domain

Customer, supplier, product and location identity behave differently.

A person may change address and surname. A legal supplier may have stable regulatory identifiers. Products may require specification, manufacturer and hierarchy evidence. Locations can share names but differ by geographic coordinates or organizational purpose.

IBM MDM similarly allows different matching algorithms to be configured for different entity types.

Entity resolution is domain-specific governance implemented through matching technology—not one universal similarity algorithm.

The graph should preserve time as well as relationships

Business relationships are not permanent.

Organizations acquire and divest subsidiaries. Suppliers change parent companies. Customers move between account structures. Products change classifications.

For many analytical, regulatory and AI use cases, the enterprise therefore needs to know not only whether a relationship exists, but when it was valid.

A relationship without effective time can give today's organizational truth to yesterday's business event.

A practical architecture: Resolve, Anchor, Relate, Qualify, Govern, Serve

I would summarize relationship-aware MDM through six steps.

1. Resolve: determine which source records represent the same real-world entity.

2. Anchor: assign or maintain a persistent enterprise identity that survives source-system change.

3. Relate: connect the entity to approved business relationships, hierarchies and relevant cross-domain context.

4. Qualify: attach authority, provenance, evidence and effective time to relationships where those distinctions matter.

5. Govern: control ambiguous identity decisions, relationship changes, merge and unmerge actions and policy exceptions.

6. Serve: expose trusted identity and contextual relationships to operational applications, analytics and AI.

Resolve → Anchor → Relate → Qualify → Govern → Serve

Do not begin with the graph database

A technology-first implementation can easily produce a sophisticated graph whose business semantics and authority are unclear.

Before choosing graph technology, the enterprise should answer more basic questions.

Which real-world entities require persistent enterprise identity?

Which relationships matter to actual business decisions?

Which relationships require authoritative governance?

Which can remain inferred or analytical?

Which history must be retained?

Which consumers need to traverse or query these relationships?

Once those requirements are clear, the storage and graph-processing architecture can be selected more rationally.

Seven questions before adding relationship-aware entity resolution

1. What real-world identity decision are we trying to improve?

2. Which attributes provide authoritative identity evidence, and which provide only similarity signals?

3. Which relationships legitimately strengthen or weaken the identity decision?

4. Can the enterprise distinguish approved relationships from observed, inferred and unresolved ones?

5. What happens to relationships when entities are merged, split or unlinked?

6. Can Stewards reconstruct why an identity or relationship decision was made?

7. Does the business problem actually require graph-native technology, or can the existing MDM relationship model satisfy the requirement?

If these answers are unclear, adding more graph technology may increase complexity before it increases trust.

The objective is not zero duplicates

“Zero duplicate data” sounds like an attractive target.

In complex enterprise environments, it can also encourage dangerous behavior if teams lower the threshold for merging records simply to reduce a duplicate metric.

Some ambiguity is legitimate.

Two records may remain unresolved because available evidence is insufficient. That is often safer than forcing them into one entity.

An unresolved identity is visible uncertainty. A false merge is hidden certainty.

Relationship-aware identity becomes increasingly important for AI

Enterprise AI needs more than a customer or supplier record.

It may need to understand ownership, hierarchy, product relationships, locations, account structures and other business connections before acting.

SAP's current Business Data Cloud architecture explicitly uses knowledge graphs to map relationships across entities, data and processes so applications and AI agents can reason with richer business context.

SAP Reltio similarly exposes resolved multidomain entities together with relationships, hierarchies and interactions for applications and AI agents.

AI does not make identity resolution less important. It increases the cost of getting identity and relationships wrong at machine speed.

Trusted identity comes before contextual intelligence

Knowledge graphs can dramatically improve how enterprises represent context.

But a graph is most valuable after the enterprise has established what its nodes represent and how strongly each relationship should be trusted.

That sequence matters.

First resolve the entity. Then preserve its identity. Then connect it to governed context. Then make that context reusable.

Knowledge graphs do not replace Master Data Management. They extend trusted identity into trusted context.


Sources & Further Reading

IBM — Master Data Management
Current IBM product material describing machine-learning-based entity resolution, multidomain master data and relationship discovery as complementary MDM capabilities.
IBM

IBM Documentation — Matching Your Data to Create Master Data Entities
Current IBM documentation describing how matching is used for deduplication and entity resolution and how matched source records are collected into master-data entities.
IBM Documentation

IBM Documentation — Matching Algorithms in IBM Master Data Management
Current IBM documentation covering standardization, bucketing, comparison, entity-specific algorithms and match-threshold configuration.
IBM Documentation

IBM Documentation — Configuring a Potential Match Workflow
IBM documentation describing how uncertain matching cases can be converted into Steward tasks for explicit match or no-match decisions rather than being forced into automatic resolution.
IBM Documentation

IBM Documentation — Discovering Relationships in Your Master Data
Current IBM documentation covering rule-based relationship discovery across records and entities, including provenance metadata identifying relationships created through discovery rules.
IBM Documentation

SAP — Master Data Management in SAP Business Data Cloud
Current SAP material describing entity resolution, golden records, relationships and dynamic knowledge graphs as connected parts of the MDM context delivered to applications and AI agents.
SAP

SAP — Business Data Fabric in SAP Business Data Cloud
Current SAP architecture describing governed semantic models and knowledge graphs that map relationships across business data, processes and entities to support contextual understanding.
SAP

SAP — SAP Reltio in SAP Business Data Cloud
Current SAP material describing cloud-native multidomain MDM with entity resolution, intelligent data graphs, relationships, hierarchies and governed entity views for operational applications and AI agents.
SAP

SAP — What Is SAP Business Data Cloud?
Current SAP architecture describing the knowledge core, business semantics and SAP Knowledge Graph as mechanisms for connecting enterprise entities, processes and data for analytics and AI.
SAP


Method Note

This article distinguishes matching, entity resolution and knowledge-graph / relationship capabilities for analytical clarity. Individual MDM products may combine these capabilities inside one platform and may use different terminology.

The framework Resolve → Anchor → Relate → Qualify → Govern → Serve and the statement that a knowledge graph represents the governed neighborhood around a resolved entity are practitioner concepts developed for this article rather than formal vendor reference architectures.

This article deliberately avoids universal enterprise duplicate-rate estimates, fixed match-accuracy claims, assumed cost savings, fixed automatic-merge thresholds and standardized implementation timelines. Appropriate matching policy depends on domain, source-data quality, available identifiers, false-match consequences, false-nonmatch consequences and stewardship capacity.

Relationship information is treated as potential identity evidence as well as downstream business context. The article does not claim that graph structure alone is sufficient to resolve identity or that graph-based technology will always outperform appropriately designed attribute-based matching.

The term Knowledge Graph is used primarily as a logical representation of entities, semantics and governed relationships. A dedicated graph database may be appropriate for some architectures but is not assumed to be mandatory for relationship-aware MDM.

Reviewed: October 2026

Previous: MDM #4. Self-Healing Master Data: From Error Detection to Controlled Recovery

Next: MDM #6. Why Digital Transformation Initiatives Struggle — The Hidden Role of Master Data

This article is part of the Enterprise MDM Series exploring how master data is evolving for AI-ready and agentic enterprises.

Comments

Popular posts from this blog

How to Read the 2026 Gartner MDM Magic Quadrant: Five Leaders, Two Acquisitions, One Buyer Framework

AI Agents: Chatbots, RPA and Agentic AI Explained

Why Enterprise MDM Governance Fails After Go-Live — and How to Make Ownership Real