The taxonomy of knowledge graphvalidation approaches

Abstract

Relevance. Knowledge graphs encoded in Resource Description Framework have become a standard foundation for data integration, semantic search, and automated reasoning across domains ranging from biomedicine to enterprise knowledge management. As these graphs grow in scale and are maintained by distributed teams, it becomes increasingly important to ensure that their data follows the defined structural and semantic rules. Aim. This survey presents a systematic review of knowledge graph validation approaches published between 2011 and 2026, covering three principal families of methods and proposes a taxonomy that classifies existing papers within following three families: (i) ontology-based validation via Web Ontology Language reasoning, (ii) constraint-based validation using Shapes Constraint Language, Shape Expressions Language, and SPARQL Inferencing Notation (SPIN) language, and (iii) statistical and machine-learning-based error detection using knowledge graph embeddings and contrastive learning. The following objectives were conducted: to review the literature sources, to carry out comparative analysis across listed above methods, to suggest taxonomy of graph validation approaches, to discover gaps in research field. Scientific novelty: the survey introduces a single three-family taxonomy that characterizes each family by its underlying assumptions (Open World Assumption vs. Closed World Assumption), expressiveness, and computational cost, and it formally defines constraint drift, the progressive misalignment of Shapes Constraint Language shapes with their underlying ontology. Practical significance. The survey provides architectural design choices that resolve the tension between the Open World Assumption (OWA) of Web Ontology Language reasoning and the Closed World Assumption (CWA) of Shapes Constraint Language shapes across all three families, serving as a reference for practitioners and a roadmap for researchers. Results: a taxonomy of validation approaches is proposed and a comparative analysis of the three families is conducted. Constraint-based validation and Shapes Constraint Language in particular, is identified as the practical standard for Resource Description Framework data quality enforcement, with ontology-based reasoning complementing it for logical consistency and machine-learning methods extending it to distributional anomaly detection. Four open research gaps are identified: provenance-aware validation, incremental validation, validation under ontology evolution (including constraint drift, formally defined in subsection Gap 3), and hybrid constraint-plus-machine-learning approaches. Conclusions: no single family satisfies all validation requirements; ontology-based, constraint-based, and statistical/ML-based methods are complementary, and the OWA/CWA inconsistency remains a central design consideration across all three. Closing the identified gaps, in particular constraint drift and hybrid approaches, defines a roadmap for future work requiring coordinated effort across the standardisation, knowledge-graph-engineering, and machine-learning-on-graphs communities.