Knowledge Graph Framework Automates Multimodal Data Validation Across Federated Data Spaces

Decentralized data sharing environments, widely conceptualized as data spaces, allow autonomous organizations to exchange digital assets directly without surrendering control to centralized repositories. Within these federated ecosystems, robust computational data governance mechanisms are essential to guarantee compliance with organizational agreements, statutory mandates, and technical data standards. While current governance infrastructures successfully enforce data access control by establishing which entities may read assets and data usage rights by regulating permissible analytical operations, they lack automated mechanisms for verifying whether data assets actually satisfy specific quality and compliance criteria. To address this structural shortfall, a research team led by Achraf Hmimou, Petar Jovanovic, Sergi Nadal, Oscar Romero, and Anna Queralt has designed and demonstrated a knowledge graph-based framework that operationalizes and automates data validation inline with data policies. Published in Data Science and Engineering, their approach introduces an intermediate semantic abstraction termed policy checkers, which represent technology-agnostic data validation plans derived directly from high-level federation policies and subsequently compiled into executable, technology-specific user defined functions.

Contemporary data management in enterprise and scientific domains increasingly relies on collaborative data analytics, big data infrastructures, and machine learning pipelines that span institutional borders. To reconcile collaborative utility with strict privacy and security standards, organizations are progressively moving away from centralized platforms in favor of federated data ecosystems. In a standard data space, trusted participants operate as independent data providers and data consumers, engaging in direct data exchange without relying on intermediate storage hosts. This architectural model promotes data sovereignty and provenance control, yet it shifts the primary governance burden to a designated federator role responsible for designing and deploying compliance protocols. The federator must implement mechanisms that resolve syntactic and semantic interoperability hurdles while verifying adherence to complex regulations such as the General Data Protection Regulation and the FAIR data principles of findability, accessibility, interoperability, and reusability. Without automated validation routines capable of scrutinizing underlying records prior to analytics, data spaces remain vulnerable to silent data corruption, structural schema drift, privacy leaks, and algorithmic bias.

The foundational architecture developed by the authors organizes federated governance across three complementary structural tiers: the Data Product Layer, the Data Platform Layer, and the Federated Computational Governance Layer. Within the Data Product Layer, all shared assets, including tabular files, medical images, and trained machine learning models, are formalized as distinct logical units termed data products. Each registered data product incorporates specific structural metadata, encompassing an attribute profile, a dataset type template specifying the underlying format, technology aspect descriptors indicating local access paths or application programming interfaces, and a formal data contract. These data contracts link localized product attributes to global common data model features through Global-Local-as-View schema mappings while formalizing binding governance policies. Above this tier, the Data Platform Layer functions as the operational gateway handling asset registration and routing consumer analytical requests. Crucially, the overarching Federated Computational Governance Layer establishes guidelines, maintains metadata integrity, and manages a comprehensive Semantic Data Model that acts as the semantic backbone encoding all system annotations.

To formalize governance agreements across disparate institutions, data spaces frequently utilize open semantic standards such as the Open Digital Rights Language. An Open Digital Rights Language policy is structurally composed of rules detailing permissions, prohibitions, or duties assigned to distinct parties regarding defined assets, with specific operational conditions refined through constraint expressions comprising operators and operands. While these open vocabularies provide expressive vocabularies capable of modeling business terms, legal constraints, and data privacy vocabulary concepts, they fundamentally lack formal mathematical semantics. Because such vocabularies cannot capture the exact algorithmic behaviors required for programmatic execution, existing policy engines cannot translate them into executable verification scripts without additional structural context. This semantic ambiguity frequently results in inconsistent policy evaluations across differing organizational connectors, impedes cross-organizational interoperability, and prevents data spaces from applying automated reasoning engines to enforce data quality dimensions, including schema validation, consistency rules, fairness metrics, and statistical properties.

To bridge this divide between high-level policy declarations and concrete execution environments, the authors structured their framework around a decoupled, two-stage data validation workflow orchestrated by two primary functional components: the PLANNER and the EXECUTOR. When an external data consumer submits an analytical service request targeting a common data model, the system queries the federated catalog to identify all mapped candidate data products. The PLANNER component then ingests the semantic metadata associated with each target product alongside a curated library of graph transformation rules provided by the federated governance team. By executing these graph transformation rules over the semantic data model, the PLANNER generates a customized collection of policy checkers. Each policy checker functions as an intermediate, technology-agnostic validation plan composed of discrete logical operations, input and output definitions, and an eventual compliance report node, translating abstract policy constraints into explicit validation logic without tying the operations to a concrete programming language or storage engine.

Generating consistent policy checkers requires resolving both syntactic dependencies and semantic policy patterns without introducing circular transformations or execution conflicts. To govern this sequence, the authors designed a formal graph transformation process using paired morphisms that detect matching subgraphs on their left-hand side and substitute them with updated structures on their right-hand side. The framework arranges these transformation rules into a strict two-layer execution hierarchy. Layer 1 executes data dependency rules, inspecting the technology aspects of the target data product to establish data access paths, identify format specifications, and generate initial data loading operations. Layer 2 subsequently processes policy pattern rules, converting abstract policy elements, such as null-value prohibitions or cryptographic anonymization mandates, into concrete operational nodes linked to the loaded data streams. Rules within Layer 2 are executed only after all Layer 1 dependencies are fully resolved, ensuring deterministic plan generation before a boolean compliance report node is appended to the policy checker graph.

Once the PLANNER produces the intermediate policy checker graphs, the workflow activates the EXECUTOR component to generate and run concrete validation scripts. The EXECUTOR relies on an extensible repository of code metadata that maps abstract operations and dataset type templates to concrete programming implementations, required library dependencies, and parameter configurations. Utilizing dynamic code generation and templating engines, the EXECUTOR resolves dependencies between the policy checker nodes and code metadata records, populating standardized source templates with targeted evaluation algorithms. The generated code is dynamically compiled into the local runtime environment as specialized user defined functions, utilizing function decorators to orchestrate automated pre-processing and post-processing stages. These user defined functions directly interrogate the physical data assets in their native formats, verify compliance against the codified rules, and write intermediate algorithmic metadata alongside final boolean compliance determinations directly back into the semantic data model as assertional annotations.

The research team evaluated the technical feasibility and computational characteristics of their framework using a fully functional prototype implemented in Python, leveraging the RDFlib library for knowledge graph parsing and SPARQL query processing. Experimental validation was conducted on a host system equipped with an eight-core Intel i5-8350U central processing unit running at 3.6 gigahertz and 16 gigabytes of random-access memory. To rigorously evaluate the framework on multimodal data structures representative of complex data spaces, the authors incorporated the UPENN-GBM clinical research dataset, which combines tabular comma-separated values containing patient demographics with complex medical imaging files encoded in the Digital Imaging and Communications in Medicine format. The prototype was assessed against five established design science testable propositions for data quality tools in data spaces: simplicity and clarity of reporting, fitness for use across varied validation tasks, processing efficiency, and connectivity with standard data space connector architectures.

To quantify processing efficiency and horizontal scalability, the authors executed comparative benchmarks contrasting their user defined function framework against RDFValidator, a conventional baseline that transforms raw datasets into Resource Description Framework triples and executes validation via the Shapes Constraint Language. When processing tabular records scaled exponentially from an initial baseline of 672 rows, RDFValidator exhibited steep exponential execution times, requiring 793.93 seconds to process 343,552 records before failing completely on larger volumes, with shape constraint evaluations accounting for roughly 75 percent of the total runtime. In stark contrast, the authors’ framework validated the identical 343,552 rows in just 0.44 seconds and successfully scaled to 21,987,328 rows in 27.5 seconds. Through interpolation, the authors estimated that evaluating 21,987,328 rows via the Resource Description Framework pipeline would consume approximately 65,323.05 seconds, or over 18 hours. While graph transformations in the PLANNER introduce non-linear overhead due to the NP-complete nature of SPARQL subgraph pattern matching, direct execution over native formats entirely eliminates data conversion bottlenecks.

The study concluded by validating architectural connectivity and integration paths with two prominent open-source data space infrastructures: Eclipse DataSpace Components and the FIWARE Data Space Connector. In an Eclipse DataSpace Components deployment, the authors mapped the Semantic Data Model into the Federated Catalog, positioning the federated team as a governance authority that synthesizes policy functions on demand for the platform Policy Engine. For FIWARE environments, the team outlined modifications to the ODRL-PAP access and usage service, extending the Policy Enforcement Point and incorporating a Data Validation Agent within the Policy Decision Point to perform dual-phase policy verification. The authors emphasize that while their prototype successfully validated privacy metrics, statistical boundaries, schema constraints, fairness indicators, and consistency rules across static multimodal assets, real-world deployment requires ongoing research. Future extensions must address automated quality requirements engineering, streaming data validation, performance optimizations for complex graph rewrite sequences, and the automated application of remedial data transformations during ongoing data product lifecycles.

Subject of Research: Automated data validation and computational governance frameworks in federated data spaces

Article Title: Operationalizing and Automating Data Validation in Data Spaces

Article References: Hmimou, A., Jovanovic, P., Nadal, S., Romero, O., & Queralt, A. (2026). Operationalizing and Automating Data Validation in Data Spaces. Data Science and Engineering, 11(2), 303-323. https://doi.org/10.1007/s41019-025-00317-7

Image Credits: AI Generated

DOI: 10.1007/s41019-025-00317-7

Keywords: data spaces, data validation, knowledge graphs, computational governance, policy checkers, user defined functions, ODRL, SHACL, multimodal data, federated data systems, Eclipse DataSpace Components, FIWARE, UPENN-GBM dataset, dynamic code generation, SPARQL, data quality