Matt Aslett's Analyst Perspectives

I recently described how business data catalogs are evolving into data intelligence catalogs. These catalogs combine technical and business metadata and data governance capabilities with knowledge graph functionality to deliver a holistic, business-level view of data production and consumption. The concept of the knowledge graph has been part of the data sector for decades, but adoption has typically been limited to industries and enterprises focused on the Semantic Web, such as media, publishing, life sciences and pharmaceuticals. However, it is worth taking a closer look at knowledge graphs. Not only do they promise to surface the relationships between entities and associated data to enable enterprises to understand how, when and where data is used across the organization, but also to enhance enterprise applications driven by generative artificial intelligence.

A knowledge graph is a structured representation of information that identifies entities, their attributes and the relationships between them. In the context of information management, the term knowledge graph emerged in the 1970s, but just as with graph databases, the concept relies on graph theory, the origins of which date back to 1736. Knowledge graphs have historically been closely associated with graph databases designed to natively store data as entities and attributes, as well as relationships. As such, many examples of knowledge graph implementations are in industries that have embraced the Resource Description Framework graph model used in Semantic Web applications. Key industries include media, publishing, life sciences and pharmaceuticals.

However, a knowledge graph does not require that data and information be persisted in a graph database. It also provides an abstraction layer that represents the relationships between distributed and disparate information sources. Search engines are a primary example of a use case that benefits from a knowledge graph without requiring the underlying data and information to be persisted in a graph database. Social networking sites such as LinkedIn and information sources such as Wikipedia also benefit from the identification of relationships between entities and attributes provided by knowledge graphs.

From an enterprise perspective, the adoption of knowledge graphs has been boosted by the emergence and evolution of the market for data catalog products. One of the primary design goals for data catalogs was to bring the self-service discovery benefits of search engines to enterprises’ data. Metadata-based data inventory is a core component of data catalogs that surfaces information from the underlying data stores to enable search-based discovery. Knowledge graph capabilities implemented by data catalog providers facilitate search-based data discovery by identifying, classifying and maintaining a map of relationships between data assets to provide additional value. For example, knowledge graph functionality helps identify the dependencies between business intelligence reports and dashboards and the complex web of data pipelines that transform and integrate the data on which they depend.

Knowledge graphs depend on multiple elements, including ontologies and semantic models that provide formally agreed definitions and meaning related to concepts, terms and relationships. As such, creating and maintaining knowledge graphs can be highly complex and manual, requiring collaboration between data and information management specialists and subject matter experts. While creating and maintaining knowledge graphs remains complex, the implementation of artificial intelligence models that take advantage of metadata surfaced by data catalogs to automatically create an ontology representing the relationships in the data reduces manual effort.

We also see knowledge graphs as critical to the delivery of data as a product by providing a representation of data and metadata usage and the relationships between data elements. In combination with curated semantic data definitions that provide a common understanding of the data, knowledge graphs enable enterprises to understand how data ownership maps to logical business units and organizational structures. Like the adoption of data as a product, the incorporation of knowledge graph capabilities in data intelligence products remains nascent. ISG’s Buyers Guide for Data Intelligence highlighted that only 44% of software providers evaluated were graded at A- or above for the incorporation of knowledge graph to map the relationships between data and entities. However, I assert that through 2027, data catalog providers will evolve their products to support data intelligence by prioritizing delivery of knowledge graph and data product platform capabilities, as well as the use of AI.

In addition to benefiting from AI, knowledge graphs also benefit enterprise AI projects. I previously discussed the role that retrieval-augmented generation plays in improving enterprise trust in GenAI applications by augmenting foundation models with real-life data and context from enterprise information. RAG is achieved by creating vector embeddings—multi-dimensional mathematical representations of entities and attributes—used to enable the identification and retrieval of similar or related data. Add a knowledge graph to the equation via an approach known as GraphRAG, and enterprises augment foundation models with information based not only on the semantic similarity between entities and attributes but also on logical connections between them. Proponents argue that the addition of information related to relationships and dependencies between entities and attributes can improve GenAI model accuracy and reasoning compared to vector-based RAG.

Creating and maintaining a knowledge graph has historically been a significant undertaking. The level of manual effort required can be alleviated by AI, while we see knowledge -graph adoption as beneficial to AI workloads. Since AI is a forcing function for enterprises to take steps to improve data and analytics processes to be more agile, adaptive and automated, I recommend that enterprises evaluate the potential benefits of knowledge graphs and include an assessment of knowledge-graph functionality in evaluations of data catalog products.

Regards,

Matt Aslett