Knowledge becomes much more useful when the system can represent not only what things are called, but how they are connected.
A person works for an organisation. A book was written by an author. A station belongs to a rail line. A chemical has a molecular formula. A school is located in a district. A historical event happened before another event and involved several entities.
Knowledge graphs represent this world as entities and explicit relationships that can be traversed. The graph does not merely store isolated facts. It stores enough structure that a system can move from one fact to related facts through typed connections.
This article sits beneath How Knowledge Works, How Metadata Works, How Networks Work and the previous How Entity Resolution Works. Networks owns graph structure broadly. Metadata owns description. Entity Resolution owns identity reconciliation. Knowledge graphs combine those layers into a traversable semantic representation.
1. Nodes Represent Entities or Concepts
A node can represent a person, place, organisation, document, event, concept, product, biological entity or other defined object.
The useful part is that the node has identity beyond its display label. “Singapore” can be shown in several languages while the graph still points to one resolved entity.
2. Edges Carry Typed Relationships
Connections are meaningful because they have a type.
“Author A — wrote → Book B” is different from “Author A — reviewed → Book B.” “Station X — part of → Line Y” differs from “Station X — located in → District Z.”
The edge label turns a generic connection into a claim that can support reasoning.
3. Properties Add Descriptive State
Nodes and edges can also carry properties: dates, names, coordinates, identifiers, status, confidence or provenance.
This lets the graph preserve both relational structure and local descriptive state.
4. Ontologies or Schemas Define What Relations Mean
A graph becomes more coherent when the allowed entity types and relationship types are defined explicitly.
The schema may state that an author is a person, that a book is a creative work, and that “wrote” links a person to a work.
The system can remain flexible without becoming semantically arbitrary.
5. Entity Resolution Prevents Fragmented Graphs
If “NUS,” “National University of Singapore” and “National Univ. of Singapore” become three unrelated nodes, the graph fragments one institution into three shadows.
Entity resolution reconciles those records so relationships accumulate around one defensible identity.
6. Controlled Vocabularies Stabilise Relationship Language
One dataset may use “located_in,” another “is in,” another “place.” Without semantic alignment, the same relation can fragment across incompatible labels.
The previous How Controlled Vocabularies Work article explains the shared-term layer that helps keep graph semantics coherent.
7. Traversal Is the Core Superpower
A table query can retrieve records matching filters. A graph query can follow relationships.
Ask: Which books were written by authors who studied at universities in Singapore? Which MRT stations connect to lines serving hospitals? Which historical events involved institutions founded before a given date?
Traversal lets the system answer multi-hop questions whose structure is naturally relational.
8. Multi-Hop Reasoning Needs Restraint
More hops do not automatically mean more knowledge.
Each edge carries uncertainty, scope and provenance. A five-hop chain can combine several weak or context-bound claims into one apparently strong conclusion.
Graph traversal should therefore preserve evidence lineage and relationship confidence rather than treating every edge as equally certain.
9. Provenance Turns a Triple Into a Defensible Claim
“A — founded → B” is useful. It becomes more trustworthy when the graph can also say where the claim came from, when it applied and whether the source is primary, secondary or inferred.
Knowledge graphs should not erase the evidence behind the edges. The graph is a representation of claims, not reality speaking without mediation.
10. Time Makes Relationships Stateful
A person can work for one organisation in 2010 and another in 2026. A station can change line ownership. A country can change name or boundary.
If time is omitted, old and current relationships can collapse into one misleading timeless graph.
Mature knowledge graphs therefore attach temporal scope where the relation changes over time.
11. Graphs Can Connect Heterogeneous Data
One source is a database table, another a document collection, another a spatial dataset, another a catalogue. A graph layer can connect entities across those systems without requiring every source to become one physical database.
The graph becomes an integration layer above distributed owners.
12. Worked Example: Library Knowledge Graph
A library graph can connect authors, works, editions, publishers, subjects, places and citations.
A reader can move from one book to its author, from the author to related works, from a work to subjects, then from those subjects to other authors.
The library-specific owners remain How a Library Works and How Library Catalogues Work.
13. Worked Example: MRT Knowledge Graph
A graph can connect station, line, asset, maintenance event, incident, supplier, component type and location.
When one component family develops a fault, the system can traverse from component type to installed assets, from assets to stations, and from stations to service consequence.
The railway-specific owner remains How MRT Works | It’s Mathematics.
14. Worked Example: Education Knowledge Graph
A curriculum graph can connect concepts, prerequisites, examples, assessment items and misconceptions.
If a learner misses one item, the system can traverse backward into prerequisite concepts and sideways into related representations rather than simply returning “wrong.”
The education-specific owner remains How Education Works.
15. Worked Example: Civilisation Atlas
A civilisation graph can connect places, institutions, technologies, trade routes, events, cultures and evidence sources across time.
The graph becomes powerful when it preserves uncertainty and competing interpretations instead of forcing history into one unquestioned storyline.
16. Knowledge Graphs Can Create False Authority
A clean graph visual can make uncertain claims look objective.
If the underlying data contains wrong entity merges, stale relations or unsupported edges, graph traversal spreads those errors further.
The graph’s elegance is not evidence of the graph’s truth.
17. A Knowledge-Graph Checklist
- Define entity types and stable identities.
- Define relationship types clearly.
- Preserve properties and temporal scope.
- Resolve duplicate entities cautiously.
- Use controlled vocabulary or schema for relation semantics.
- Attach provenance to claims.
- Represent uncertainty where edges are inferred or contested.
- Limit multi-hop reasoning when evidence weakens across the chain.
- Keep source systems as canonical owners where appropriate.
- Test graph queries against real receiver tasks.
18. Read the Mechanism Forward, Backward and Sideways
Forward: source records → entity resolution → typed relations → graph → traversal → answer. Backward: start from a wrong multi-hop answer and trace which edge or identity created the error. Sideways: compare librarian, database engineer, researcher and learner. The same graph can support discovery, integration, explanation and diagnosis if the semantics remain explicit.
19. The Civilisation Lesson
Civilisation becomes intelligent when records can be connected without collapsing ownership. A knowledge graph is one way to let separate facts become a navigable field while still preserving where each fact came from.
A knowledge graph is not knowledge because it has nodes and edges. It becomes knowledge infrastructure when the identities, relations and evidence behind those edges are strong enough to survive traversal.
Continue through How Entity Resolution Works, How Knowledge Works and the master How X Works hub. Next: persistent identifiers — how a knowledge object keeps identity even when its location or version changes.