Knowledge Graphs

Introduction

• graph theory by Euler

There are different graph models to choose from:

• directed vs undirected graphs (edges have direction or not)
• weighted vs unweighted graphs (edges have weights or not)
• cyclic vs acyclic graphs (graphs with cycles or not)
• hypergraphs (edges can connect more than two nodes)
• property graphs (nodes and edges can have properties)

They can all be boiled down to four basic primitives, that allows you to build any graph structure:

• nodes (entities)
• labels (names)
• relationships (edges)
• properties (attributes)

The last one can be seen in the property graph model, where nodes and edges can have key-value pairs associated with them. Is the most common model today.

Motivation

• Data needs to be understandable in order to have value
• provide contextualized understanding of data is a powerful tool to make sense of data and generate knowledge
• allows to represent complex relationships and networks very intuitively, both for humans and machines
• very flexible, just add more nodes and edges as needed

Definition

• Knowledge graphs are a specific type of graph that provide contextual understanding.
• Provides a holistic view of data, capturing not only entities but also their relationships and attributes.
• Data can come from anywhere, from self-contained graph databases, to datalakes, to federated systems.

holistic meeans dealing with or treating the whole of something or someone and not just a part. federated means a group of entities that are joined together but still maintain their own autonomy.

Organizing principle

• graphs that are functional but not understandable are not very useful
• knowledge graphs are organized around semantics
• semantics is the name for the meaning of things, so people and machines can understand them
• semantics is what makes knowledge graphs different from regular graphs
• data shouldn’t need a manual to be understood
• organizing principles are contracts between the graph and its users
• taxonomies and ontologies are high level organizing principles
• taxonomies are hierarchical classifications of concepts
• taxonomies can be defined as Category nodes connected by SUBCLASS_OF relationships
• ontologies are more complex relationships between concepts
• ontologies allows users and machines to take actions based on the data
• there are several off-the-shelf widely used ontologies for particular domains, such as SNOMED CT
• you can choose to build your own ontology or extend an existing one, depending on your domain and use case

Learning Path: Semantic Web

RDF Basics

• Triple structure: Subject-Predicate-Object
• URIs as identifiers: How resources are named
• Literals vs Resources: Data values vs things
• Serialization formats: Turtle (most readable), RDF/XML, JSON-LD
• Exercise: Describe yourself and 3 relationships in Turtle

2. Namespaces & Vocabularies

• Prefixes: Shorthand for long URIs (rdf:, rdfs:, skos:)
• Common vocabularies: Dublin Core, FOAF, Schema.org
• Exercise: Create your own namespace and vocabulary

3. RDFS Essentials

• Classes: rdfs:Class, rdfs:subClassOf
• Properties: rdfs:Property, rdfs:domain, rdfs:range
• Labels: rdfs:label, rdfs:comment
• Exercise: Model a simple domain (e.g., library with books, authors, genres)

4. SKOS for Organization

• Concepts: skos:Concept, skos:ConceptScheme
• Labels: skos:prefLabel, skos:altLabel
• Relationships: skos:broader, skos:narrower, skos:related
• Exercise: Build a 3-level taxonomy with multilingual labels

5. OWL (Optional Advanced)

• Only if needed: Property characteristics, restrictions, inference
• Start simple: Equivalence, disjointness
• Exercise: Add logical constraints to your RDFS model

Tools to Try

• Protégé: Visual ontology editor
• RDF Playground: Online Turtle editor
• SPARQL: Query language (learn after understanding triples)