Link Validity

This dimension evaluates if a relationship's predicate (the edge label) correctly and precisely describes the nature of the connection between the subject and object entities. A semantically incorrect link uses an imprecise or entirely wrong relationship type.

• Example of Correctness: (Paris, capitalOf, France) uses the precise predicate capitalOf.
• Example of Error: (Paris, locatedIn, France) is less precise, while (Paris, manufacturerOf, France) is semantically nonsensical.

Validation often involves checking against the ontology to ensure the predicate's domain and range are respected (e.g., a manufacturedBy predicate should connect a Product entity to a Company entity).

This dimension assesses the objective truth of the asserted triple, independent of its semantic form. A link can be semantically well-formed but factually wrong.

• Example of Accuracy: (Tesla Model S, manufacturedBy, Tesla, Inc.) is factually accurate.
• Example of Error: (Tesla Model S, manufacturedBy, Ford Motor Company) is factually incorrect, even though the predicate manufacturedBy is used correctly between a Vehicle and a Company.

Establishing factual accuracy requires verification against authoritative sources, ground truth, or domain expert validation. It is the core defense against the propagation of misinformation within the graph.

This dimension ensures links do not violate the logical rules and constraints defined in the knowledge graph's schema or ontology. Inconsistencies create contradictions that break automated reasoning.

Key checks include:

• Domain/Range Violations: A bornIn predicate (domain: Person, range: Location) linking a Company to a Person is invalid.
• Cardinality Violations: If a hasCEO property is defined as functional (max 1), a company node with two hasCEO links is inconsistent.
• Disjointness Violations: If Person and Organization are defined as disjoint classes, a single node cannot be an instance of both.

Tools like OWL reasoners (e.g., HermiT, Pellet) automatically detect these violations.

Many facts are only true within a specific context or timeframe. This dimension evaluates if a link is qualified with necessary contextual or temporal metadata, or if a static link has become stale.

• Temporal Example: (Angela Merkel, holdsPosition, Chancellor of Germany) was valid from 2005 to 2021 but is not valid for the current time without temporal scoping.
• Contextual Example: (Aspirin, treats, headache) is generally valid, but (Aspirin, contraindicatedFor, patient) is only valid in the context of a patient with certain medical conditions (e.g., Reye's syndrome).

Temporal Knowledge Graphs and named graphs (for context) are modeling patterns used to capture this dimensionality and prevent validity errors.

This is a foundational technical dimension ensuring that every link points to an actual, existing node in the graph. A breach of reference integrity creates a dangling link or broken reference.

• Cause: Often occurs during data deletion or flawed ETL processes where a target entity is removed but relationships pointing to it are not.
• Impact: Breaks query execution, cripples traversal-based algorithms, and undermines the graph's structural soundness.
• Validation: Simple integrity checks can scan all triples to verify that the object of every statement exists as a subject node in the dataset. Graph database systems like Neo4j or Amazon Neptune typically enforce this at the database level.

Link validity is measured using a combination of automated and manual techniques:

• Rule-Based Validation: Scripts or SPARQL queries check for schema conformance and logical consistency.
• Statistical Anomaly Detection: Machine learning models identify edges that are statistical outliers within local graph neighborhoods (e.g., a manufacturedBy link in a dense cluster of worksFor links).
• Gold Standard Benchmarking: Links are sampled and verified against a high-quality reference dataset to calculate metrics like precision and recall.
• Expert Sampling: Domain experts manually audit random or high-impact links.
• Inference Cross-Checking: Using reasoning engines to derive facts and checking if they contradict existing links.

Measures the proportion of entities (nodes) that correctly correspond to their real-world referents. A foundational metric, as invalid entities guarantee invalid links.

• Core Dependency: Link Validity assumes the entities at both ends of a relationship are accurate.
• Example: A node for "Apple Inc." must not be conflated with data about the fruit or other companies named Apple.

Evaluates whether all stated facts (triples) in the graph are logically non-contradictory and align with a verifiable ground truth. Directly interacts with Link Validity.

• Logical Contradiction: A graph stating (Person, employedBy, CompanyA) and (Person, employedBy, CompanyB) where the companies are disjoint creates a consistency violation.
• Temporal Consistency: Facts must be consistent within a given timeframe.

Ensures every relationship (edge) points to a valid, existing target entity, preventing dangling links. This is a syntactic prerequisite for semantic Link Validity.

• Broken Link: A triple (Product-123, manufacturedIn, Factory-456) where node Factory-456 does not exist violates reference integrity.
• Enforcement: Typically enforced by database foreign key constraints or ontology axioms.

Assesses the degree to which instances and relationships adhere to the constraints defined in the governing ontology or schema. Link Validity is often defined by this schema.

• Domain/Range Violations: A link (River, hasCEO, Person) violates typical ontology rules (hasCEO's domain should be Organization).
• Cardinality Violations: Asserting two birthPlace links for a single person when the schema defines a maximum cardinality of one.

A primary methodology for assessing Link Validity. Involves executing SHACL or SPIN constraints to automatically detect invalid relationships.

• Syntax: sh:path ex:locatedIn ; sh:class ex:Country .
• Business Logic: Rules can encode complex domain knowledge, e.g., "A subsidiary must be located in a country where the parent company operates."
• Tooling: Used by validation engines like TopBraid EDG, PoolParty, and Stardog.

A formal property where no set of facts or inferred conclusions violates the logical constraints (e.g., disjointness, property characteristics) of the ontology. Link Validity failures can cause logical inconsistencies.

• Disjointness: If City and Country are declared disjoint, a link (Berlin, isA, Country) is both invalid and creates a logical inconsistency.
• Inferred Inconsistency: Reasoners like HermiT or Pellet automatically detect these violations.