Do you need LLM or a Knowledge Graph?

What is a knowledge graph anyways?

A knowledge graph, also referred to as a semantic network, is a representation of interconnected real-world entities, such as objects, events, situations, or concepts, showcasing their relationships. Typically stored in a graph database, this information is visually depicted in a graph structure, hence the term knowledge "graph."

Comprised of three primary elements—nodes, edges, and labels—a knowledge graph uses nodes to represent any object, place, or person. Edges define the relationships between nodes. For instance, a node could represent a client like IBM, an agency like Ogilvy, and an edge would denote the customer relationship between IBM and Ogilvy.

In this structure, A signifies the subject, B denotes the predicate, and C represents the object.

It's important to note that the definitions of knowledge graphs vary, with ongoing debates suggesting their similarity to knowledge bases or ontologies. The term gained popularity through Google's Knowledge Graph in 2012.

Ontologies are frequently discussed concerning knowledge graphs, with ongoing debates on their differences. They primarily serve to create a formal representation of entities within the graph. While usually based on a taxonomy, they can encompass multiple taxonomies, thus maintaining a distinct definition. As both knowledge graphs and ontologies use nodes and edges and are based on RDF triples, they often resemble each other in visualizations.

For instance, an ontology might distinguish between events at a venue like Madison Square Garden, using variables such as time. It could differentiate between various events held by a sports team, like the New York Rangers, each identified by their date and time, despite being in the same location.

The Web Ontology Language (OWL), supported by the World Wide Web Consortium (W3C), stands as an example of a widely adopted ontology. This organizational structure of knowledge relies on technological infrastructure such as databases, APIs, and machine learning algorithms, aiding efficient access and processing of information for people and services.

How a knowledge graph operates

Knowledge graphs typically aggregate datasets from diverse sources with varying structures. Schemas, identities, and context collaborate to structure diverse data, providing frameworks, classifying nodes appropriately, and determining the context in which knowledge exists. This aids in distinguishing between words with multiple meanings, facilitating systems like Google's search engine to differentiate between contexts such as the brand Apple and the fruit apple.

Fueled by machine learning, knowledge graphs employ natural language processing (NLP) to create a comprehensive view of nodes, edges, and labels through semantic enrichment. This process allows them to identify objects and comprehend relationships in the data. Once completed, a knowledge graph enables question answering systems and search functionalities to retrieve and reuse comprehensive answers for specific queries, saving time for users. This methodology also finds application in business settings, streamlining data collection and integration for informed decision-making.

Integration efforts around knowledge graphs not only facilitate data connections but also enable the creation of new knowledge by establishing relationships between previously unconnected data points.

Use cases of knowledge graphs

Consumer-facing knowledge graphs like DBPedia and Wikidata, which source data from Wikipedia, and Google Knowledge Grap

utilized in search engine results, are setting user expectations in various industries.

Beyond consumer applications, knowledge graphs have significant utility in other sectors:

Retail: Supporting up-sell and cross-sell strategies by recommending products based on individual purchase behaviors and popular trends across demographic groups.

Entertainment: Leveraging AI-based recommendation engines for content platforms, aiding in suggesting new content based on online engagement behaviors.

Finance: Employed for know-your-customer (KYC) and anti-money laundering initiatives, aiding in financial crime prevention and investigation.

Healthcare: Benefiting the industry by organizing and categorizing relationships within medical research, helping in diagnoses and treatment plans tailored to individual needs.

Although this example may appear impressive, it actually underscores several limitations of knowledge graphs. Initially, a portion of the graph appears trivial. Furthermore, some labels likely contain inaccuracies (for instance, it's probable that Global Investment Inc controls Big Bucks Cafe), resulting in inaccuracies within the inferred facts. Moreover, the meaning behind the metadata labeled as "controls" remains unclear. Does it signify direct control or a franchise relationship?

Less obvious limitations involve the inherent incompleteness of the graph and a time-related issue. If Global Investment were to part ways with Big Bucks Cafe, the graph would require significant revisions.

The act of adding metadata in 2023 is akin to fitting an engine onto a horse-drawn cart and claiming it moves a bit faster.