Rule-Based Validation

Rule-based validation provides deterministic outcomes; a given data state and rule set will always produce the same validation result. This creates a fully auditable trail of compliance checks, which is critical for regulated industries and for debugging data pipelines. Unlike statistical methods, there is no probability or confidence score—a fact either passes or fails a rule.

• Example: A rule stating Employee age must be >= 18 will categorically flag any Employee entity with an age property value of 17.

This is the primary use case, enforcing the structural and logical constraints defined in the knowledge graph's ontology (e.g., OWL, SHACL). Rules codify the expected shape of the data.

Common constraint types include:

• Data Type: birthDate must be an xsd:date.
• Cardinality: A Person must have exactly one biologicalMother.
• Domain/Range: The employedBy property can only link a Person (domain) to an Organization (range).
• Disjointness: An entity cannot be both a Company and a Person.
• Value Sets: status must be in ["Active", "Inactive", "Retired"].

Rules move beyond syntax to enforce business logic and real-world semantics. They identify facts that, while syntactically valid, are logically impossible or inconsistent with other data in the graph.

Examples:

• Temporal Logic: contractEndDate must be after contractStartDate.
• Inverse Relationship Consistency: If Alice manages Bob, then Bob must have a reportsTo link to Alice.
• Transitive Closure: Ensure the partOf hierarchy is acyclic (no entity can be part of itself).
• Derived Fact Validation: If totalCost = unitPrice * quantity, validate this holds for all InvoiceLine items.

Rule sets are applied proactively within data ingestion pipelines (e.g., ETL/ELT processes) to act as a guardrail, preventing invalid or inconsistent data from entering the knowledge graph. This shifts quality left in the data lifecycle, reducing costly downstream cleanup and ensuring the graph remains a reliable source of truth for applications like Graph RAG and semantic reasoning.

• Implementation: Validation engines (e.g., SHACL processors, custom logic) run in-line with transformation jobs, rejecting or quarantining violating records.

Specialized declarative languages exist to define validation rules. The W3C-standard SHACL (Shapes Constraint Language) is the most prominent for RDF-based knowledge graphs. SHACL Shapes define the target classes and the constraints that their instances must satisfy.

Example SHACL Snippet:

turtleCopy
ex:PersonShape
a sh:NodeShape ;
sh:targetClass ex:Person ;
sh:property [
  sh:path ex:birthDate ;
  sh:datatype xsd:date ;
  sh:maxCount 1 ;
] .

Legacy systems may use SPARQL Inferencing Notation (SPIN) or proprietary rule engines in property graph databases.

Rule-based validation is complementary to statistical quality assessments like Embedding Quality checks or Anomaly Detection. Rules excel at checking known, predefined constraints, while statistical methods discover unknown patterns, outliers, and potential new rules.

• Rules: "Check for what you know is wrong."
• Statistics: "Discover what might be wrong." A robust quality regime uses both: rules ensure baseline integrity, and machine learning models monitor for drift and novel inconsistency patterns.

These rules enforce the formal logic defined in the ontology (e.g., OWL). They prevent contradictory statements and ensure the graph is logically coherent.

• Disjointness: A Person cannot also be a Company.
• Cardinality: An Employee can have exactly one employer.
• Property Characteristics: If hasParent is transitive, and A hasParent B and B hasParent C, then A must implicitly haveParent C.

Violations indicate a fundamental break in the knowledge model.

These are foundational checks on the format and structure of data values, ensuring they conform to expected patterns.

• Datatype Conformance: A birthDate property value must be a valid xsd:date.
• Regex Pattern Matching: An employeeID must match the pattern EMP-\d{6}.
• Value Range: A productRating must be an integer between 1 and 5.

These rules catch data entry errors and malformed inputs before semantic processing.

These rules validate the real-world meaning and plausibility of connections between entities, going beyond pure syntax.

• Domain/Range: The manufacturedBy property must link a Product (domain) to a Company (range).
• Anti-Patterns: A headquarteredIn relationship should not link a City to another City.
• Temporal Logic: An employee endDate cannot be earlier than their startDate.

Violations often represent factual errors or incorrect relationship assignments.

Rules that govern the canonical representation of entities and prevent duplication or ambiguity.

• Key Property Uniqueness: No two Person nodes can have the same official socialSecurityNumber.
• Functional Properties: An ISBN number uniquely identifies one Book.
• Same-As Consistency: If nodeA owl:sameAs nodeB, they must share all the same property values.

These are critical for maintaining a single source of truth and accurate entity resolution.

Domain-specific rules that encode organizational policies, compliance requirements, or operational knowledge.

• Compliance: A BankTransaction over $10,000 must have a linked verifiedIdentity entity.
• Lifecycle State: A PurchaseOrder with status APPROVED must have a value for the approvedBy property.
• Derived Facts: If a Customer has status = GOLD and orderTotal > $5000, then eligibilityTier must be set to PREMIUM.

These rules directly tie graph quality to business process integrity.

A formal property of a knowledge graph where no set of facts or inferred conclusions violates the logical constraints defined by its ontology. This is the primary goal of rule-based validation.

• Key Constraints: Includes checks for class disjointness (e.g., a Person cannot also be a Location), property domain/range violations, and cardinality restrictions (e.g., a person has exactly one biological mother).
• Validation Role: Rule-based systems are explicitly designed to detect and report logical inconsistencies, which are critical failures for reasoning systems.

The active process of ensuring all data in a knowledge graph complies with predefined logical, semantic, and data-type constraints of its schema. This is the operational execution of rule-based validation.

• Process: Involves running a constraint validation engine (e.g., using SHACL or SPIN) over the graph instance.
• Output: Produces a validation report detailing each constraint violation, including the specific node, property, and broken rule.

The degree to which instances and relationships in a knowledge graph adhere to the classes, properties, and hierarchies defined in its governing ontology. Rule-based validation is the primary technical mechanism for measuring conformance.

• Validation Targets: Rules check that nodes have correct rdf:type declarations, properties are used between appropriate classes, and mandatory properties (required fields) are present.
• Business Impact: High schema conformance is essential for reliable integration, querying, and automated reasoning.

Systems that perform logical inference and deductive closure over knowledge graphs using formal rule sets. They are closely related to, but distinct from, validation engines.

• Inference vs. Validation: A reasoner adds new facts based on rules (e.g., inferring :Bob :hasSibling :Alice from :Alice :hasSibling :Bob). A validator flags existing facts that break rules.
• Synergy: The ontology rules used for reasoning (e.g., property symmetry) are often the same rules checked for consistency during validation.

The capability to record and trace the origin, lineage, and transformations of each fact within a knowledge graph. This is critical for diagnosing the root cause of rule violations.

• Link to Validation: When a rule violation is detected, provenance metadata answers: Which source system introduced this fact? Which ETL pipeline transformed it? Which user approved it?
• Enables Correction: Effective remediation of validation errors requires understanding the data's journey to fix the process, not just the symptom.