What Are Knowledge Graphs?

Knowledge graphs are structured representations of knowledge in which entities appear as nodes, relationships between them appear as labeled edges, and the whole is organized by a schema or ontology that gives the labels defined meaning. The graph abstraction lets facts about people, places, products, genes, or equipment be stored, merged, and queried without committing in advance to a fixed relational schema, and it makes connections between records first-class objects rather than something recovered through joins. The term entered general use after Google described its own graph in 2012, but the underlying ideas descend from semantic networks, frame systems, and description logics developed in artificial intelligence research from the 1960s onward.

A knowledge graph differs from an ordinary graph database in the presence of explicit semantics. Classes, properties, hierarchies, and constraints are recorded in the graph itself, so that software can validate new assertions and infer statements that were never written down. A survey of knowledge graph foundations in ACM Computing Surveys treats this combination of data model, schema, and inference as the defining feature, and organizes the field around graph representation, deductive and inductive knowledge, construction, quality assessment, and refinement.

Data Models and Schemas

Two data models predominate. The Resource Description Framework represents every fact as a subject-predicate-object triple with globally unique identifiers, which makes graphs from different publishers mergeable by construction and supports vocabularies expressed in RDF Schema and the Web Ontology Language. Property graphs, used by systems such as Neo4j and the ISO-standardized GQL query language, instead attach key-value attributes directly to nodes and edges, trading formal semantics for modeling convenience. Schema layers range from lightweight to strict: SHACL and ShEx validate the shape of incoming data, while OWL axioms license logical entailment such as inferring that a component of a subsystem is also a component of the parent system.

Construction and Refinement

Populating a graph is usually the dominant engineering cost. Sources include relational databases mapped through declarative rules, semi-structured files, and free text processed by named entity recognition, relation extraction, and entity linking pipelines that attach mentions in documents to canonical identifiers. Because extraction is imperfect and sources disagree, refinement steps matter as much as extraction: deduplication and entity resolution merge coreferent nodes, provenance annotation records where each assertion came from, and completion methods predict plausible missing edges. Openly published graphs such as Wikidata, DBpedia, and the Unified Medical Language System serve as reference targets that give independently built graphs a shared vocabulary to link against.

Querying, Storage, and Embeddings

Querying is served by graph pattern languages, principally SPARQL 1.1 for RDF and Cypher or GQL for property graphs, both of which match subgraph patterns and support path expressions that traverse an unbounded number of edges. Meeting those queries at scale requires specialized engines, and a survey of RDF stores and SPARQL engines documents the indexing, join ordering, and distribution strategies they use on graphs of billions of triples. A complementary line of work embeds entities and relations as vectors through models such as TransE, RotatE, and graph neural networks, supporting approximate reasoning, link prediction, and integration with language models through retrieval-augmented generation.

Applications

Knowledge graphs have applications in a wide range of fields, including:

• Web search, question answering, and virtual assistants
• Product catalogs, recommendation, and supply chain traceability
• Drug discovery, clinical decision support, and biomedical data integration
• Fraud detection, anti-money-laundering, and financial risk analysis
• Enterprise data integration and master data management
• Cybersecurity threat intelligence correlation
• Digital twins and industrial asset management