Knowledge Graph

Entities are the primary objects or concepts represented in a knowledge graph. Each entity is a distinct, identifiable node, such as a person, organization, location, product, or event. Entities are the foundational subjects and objects of facts.

• Real-world representation: A node for "Marie Curie", "Polonium", or "Nobel Prize in Physics".
• Unique identifiers: Each entity is assigned a persistent URI (Uniform Resource Identifier) to ensure it can be uniquely referenced and linked across datasets.
• Types and classes: Entities are typically instances of a class or type defined in an ontology (e.g., Marie Curie is of type Scientist and Person).

Relationships are directed, labeled connections between two entities that define how they are associated. These edges form the "graph" structure and encode the semantic meaning of the connection.

• Defined semantics: A relationship like discovered connects Marie Curie to Polonium. A relationship awardedTo connects Nobel Prize in Physics to Marie Curie.
• Property graph model: In systems like Neo4j, relationships are first-class citizens that can themselves have properties (e.g., a collaboratedWith relationship could have a startYear property).
• Logical inference: Relationships enable reasoning; if Polonium isElementDiscoveredBy Marie Curie and Marie Curie hasProfession Scientist, one can infer a Scientist discovered Polonium.

An ontology is the formal schema or data model that defines the vocabulary of a knowledge graph. It specifies the allowed types of entities (classes), the types of relationships (properties), and the rules governing their use.

• Class hierarchy: Defines a taxonomy (e.g., Scientist is a subclass of Person).
• Property constraints: Dictates the domain and range of relationships (e.g., the discovered relationship can only link a Scientist (domain) to a ChemicalElement (range)).
• Reasoning rules: Enables logical consistency checks and new fact inference through formalisms like OWL (Web Ontology Language). Common ontologies include Schema.org for the web and FOAF (Friend of a Friend) for describing people.

Attributes are data values that describe properties of entities or relationships. Unlike relationships that link to other entities, attributes link to raw data values like strings, numbers, or dates.

• Entity properties: Marie Curie has an attribute birthDate with the literal value "1867-11-07". Polonium has an attribute atomicNumber with value 84.
• Data typing: Literals are typed (e.g., xsd:date, xsd:integer, xsd:string) to ensure proper interpretation and enable range queries.
• Distinction from relationships: A key design choice is whether to model a piece of information as an attribute (literal) or a relationship to a new entity (e.g., birthPlace as a string "Warsaw" vs. a relationship to an entity Warsaw of type City).

The triple store is the foundational storage and query layer for a knowledge graph, based on the Resource Description Framework (RDF). Data is stored as subject-predicate-object triples, which directly map to entity-relationship-entity or entity-attribute-value statements.

• Atomic fact representation: The triple <Marie_Curie> <discovered> <Polonium> is a single, atomic unit of storage.
• SPARQL querying: The SPARQL protocol and query language is the standard for retrieving and manipulating data from an RDF triple store. It allows for complex graph pattern matching and federated queries across multiple graphs.
• Named Graphs: Triples can be grouped into named sub-graphs, allowing for provenance tracking, access control, and versioning of specific subsets of knowledge.

Entity Resolution is the critical process of determining when references in different data sources or within the same graph refer to the same real-world entity. It ensures the "singleness" of an entity node, preventing duplication and enabling data fusion.

• Same-As Links: Uses owl:sameAs or skos:exactMatch predicates to assert that two entity URIs are identical (e.g., linking a DBpedia URI for Marie Curie to a Wikidata URI for the same person).
• Deduplication algorithms: Employ probabilistic matching, rule-based systems, or machine learning models to compare entity attributes and relationships.
• Foundation for linking: This process is essential for creating a Linked Data web, where knowledge graphs published by different organizations are interlinked, forming a global, decentralized graph of knowledge.

Knowledge graphs power semantic search by understanding user intent and the relationships between concepts, moving beyond keyword matching. They enable:

• Faceted navigation and dynamic filtering based on entity properties.
• Disambiguation of terms (e.g., 'Apple' the company vs. the fruit).
• Intelligent recommendations by traversing connections between products, research papers, or internal documents.

Example: A pharmaceutical company uses a knowledge graph to connect drug compounds, target proteins, clinical trials, and research authors, allowing scientists to discover relevant literature and potential collaborators in seconds.

The explicit relationships in a knowledge graph allow AI systems to perform logical inference and multi-hop reasoning. This is critical for:

• Question Answering: Answering complex queries like "Which drugs target protein X and have passed Phase 2 trials?" requires chaining multiple facts.
• Hypothesis Generation: Identifying novel connections, such as potential drug repurposing opportunities by linking disease pathways.
• Consistency Checking: Detecting contradictions in enterprise data (e.g., a supplier listed as 'inactive' but still receiving new POs).

This capability provides a deterministic factual grounding that complements the probabilistic nature of LLMs.

Knowledge graphs act as a unified semantic layer over disparate, siloed data sources (databases, CRMs, ERPs). They provide:

• A canonical model (ontology) that defines a single source of truth for core entities like 'Customer' or 'Product'.
• Entity resolution to deduplicate and link records referring to the same real-world object across systems.
• Data lineage by explicitly modeling how data flows and transforms between systems.

This application is foundational for breaking down data silos and enabling a 360-degree view of customers, supply chains, or research domains.

When integrated with RAG architectures, knowledge graphs move retrieval beyond simple semantic similarity to graph-augmented retrieval. This involves:

• Using the graph to retrieve not just a single relevant chunk, but a subgraph of connected facts to provide full context to the LLM.
• Traversing relationships to gather supporting evidence from multiple, connected documents.
• Filtering and prioritizing retrieved information based on entity authority or recency encoded in the graph.

This results in more accurate, factually consistent, and less hallucinatory LLM outputs by providing structured, verifiable context.

Unlike collaborative filtering, graph-based recommendations leverage the rich relational structure between users, items, and attributes. Key techniques include:

• Graph Neural Networks (GNNs) that learn from the connectivity patterns to generate embeddings for nodes, capturing nuanced preferences.
• Path-based reasoning, such as recommending a product because a similar user liked a product from the same brand in the same category.
• Real-time personalization by updating user-node interactions and immediately influencing traversal paths.

This is used in e-commerce ("customers who viewed this also viewed"), content platforms, and B2B service recommendations.

In regulated industries, knowledge graphs model complex regulatory frameworks, organizational structures, and transaction flows to automate oversight. Applications include:

• Anti-Money Laundering (AML): Identifying suspicious transaction networks by mapping entities (people, companies, accounts) and the money flow between them to detect unusual patterns.
• Supply Chain Due Diligence: Mapping multi-tier supplier relationships to assess concentration risk, geopolitical exposure, or compliance with environmental regulations.
• Regulatory Knowledge Management: Structuring thousands of interconnected rules from bodies like the SEC or FDA, enabling automated compliance checking against internal process data.

This transforms manual, document-based audits into proactive, graph-driven monitoring.

A formal, explicit specification of a shared conceptualization. It defines the types, properties, and interrelationships of entities within a specific domain, serving as the schema or data model for a knowledge graph.

• Purpose: Provides the logical framework that dictates what can be represented and how entities relate.
• Example: A biomedical ontology defines that a 'Gene' is a type of 'BiologicalEntity', can 'encode' a 'Protein', and has properties like 'sequence' and 'location'.
• Role in KG: Without a well-defined ontology, a knowledge graph becomes an unstructured collection of facts, limiting its reasoning capabilities.

A standard model for data interchange on the web, representing information as subject-predicate-object triples. It forms the foundational data model for the semantic web and many knowledge graphs.

• Structure: Each fact is a triple (e.g., <Paris> <isCapitalOf> <France>).
• Flexibility: Uses URIs for unambiguous identification and can integrate data from diverse sources.
• Storage: RDF data is stored in specialized databases called triplestores or RDF stores.

The SPARQL Protocol and RDF Query Language, the standard query language for RDF data. It allows for complex graph pattern matching, aggregation, and reasoning over knowledge graphs.

• Function: Enables queries that traverse relationships, such as "Find all drugs that treat diseases caused by mutations in Gene X."
• Capability: More powerful than simple lookups, supporting joins across disparate parts of the graph and inferencing based on ontological rules.
• Analogy: Serves a similar purpose for RDF graphs as SQL does for relational databases.

An alternative graph data model where nodes (entities) and edges (relationships) can have associated properties (key-value pairs). This model prioritizes efficient traversal and is dominant in graph databases like Neo4j.

• Key Difference from RDF: Properties are attached directly to nodes/edges, whereas in pure RDF, attributes are modeled as separate triples.
• Use Case: Excellent for operational queries, fraud detection, and network analysis where traversal speed is critical.
• Interoperability: Many systems can convert between property graph and RDF representations.

The process of determining when two or more records in a knowledge graph refer to the same real-world entity. It is a critical data cleaning and integration step.

• Challenge: The same person may appear as "J. Smith," "John Smith," and "Smith, John" across different sources.
• Techniques: Uses algorithms based on similarity metrics, graph embeddings, and declarative rules to cluster duplicate references.
• Impact: Failure to perform accurate entity resolution leads to a fractured, inconsistent knowledge graph.

A class of deep learning models designed to perform inference on graph-structured data. GNNs learn representations of nodes, edges, and entire graphs by aggregating information from a node's local neighborhood.

• Application to KGs: Used for link prediction (inferring missing relationships), node classification (categorizing entities), and graph embedding.
• Mechanism: Operates via a message-passing paradigm, where nodes iteratively update their embeddings based on their neighbors' features.
• Value: Enables machine learning directly on the knowledge graph's structure, uncovering latent patterns.