A machine-learning model can estimate that two things are related. A knowledge graph can state the relationship, identify both things, connect them to other sources, define the vocabulary used and preserve where the statement came from. The two approaches answer different parts of the intelligence problem—and their combination is one of the most powerful ideas in modern AI.
01 The Web gave every document an address. Semantic technologies give every statement a structure.
The original Web created a universal architecture for publishing and linking documents. Semantic Web technologies extend that architecture so that data can also be identified, linked and interpreted consistently by software. The key move is simple and profound: relationships become first-class objects with explicit meaning.
Instead of storing a name in one table, a role in another and a programme in a third system, a knowledge graph expresses connected statements. Each subject, relationship and object can have a stable Web identifier. Other organisations can refer to the same entities without copying an entire database or agreeing on one central platform.
These are human-readable statements and machine-processable graph edges at the same time. Their value grows as they connect to vocabularies, datasets and services beyond the system where they were first written.
02 Knowledge graphs and their schemata are shared models of a domain
Knowledge graphs organise entities and the relationships between them. Their strength lies in the combination of graph structure, globally reusable identifiers and formal vocabularies. A graph can join information from research repositories, enterprise systems, cultural collections, scientific instruments and public datasets while preserving the distinctions each source needs.
Identify
Use stable IRIs so that a person, organisation, dataset or concept can be referenced unambiguously across systems.
Relate
Express typed links: authored, teaches, derived from, located in, governed by, compatible with or part of.
Interpret
Connect the graph to vocabularies and ontologies that define classes, properties, constraints and logical consequences.
This architecture supports interoperability without forcing every participant into one software product or one database schema. It is particularly valuable when data must cross institutional, disciplinary or national boundaries and remain usable over time.
03 Open standards turn one graph into an ecosystem
Fabien's course is built around W3C standards because interoperability is a property of the whole ecosystem. Each layer has a precise role.
Represents information as subject–predicate–object triples and combines those triples into graphs.
Read the W3C specificationDescribes structural expectations and validates RDF graphs against explicit shapes and constraints.
Read the W3C specificationQueries graph patterns precisely across RDF datasets, including optional paths, filters and aggregations.
Read the W3C specificationDefine vocabularies and ontologies, organise classes and properties, and support formal inference.
Explore OWL 2Publishes and links taxonomies, thesauri and classification schemes as machine-readable knowledge organisation systems.
Explore SKOSDescribe provenance, catalogues, datasets and the links between distributed data collections.
Explore provenance“He who controls metadata controls the Web and, through the Web, many things of our world.”Pr Fabien Gandon — from his official homepage
04 Reasoning makes implicit knowledge explicit
Formal semantics allows software to derive a statement from other statements following an explicit reasoning path. The result is different from a statistical guess: the conclusion follows from declared relationships and rules.
A small semantic laboratory
Follow the same information through description, validation and inference. The example is deliberately compact; the principles scale to large, distributed knowledge graphs.
SPARQL works directly with the graph structure. A query describes the pattern sought rather than the tables to join.
PREFIX dsti: <https://dsti.school/id/>
SELECT ?programme
WHERE {
?course dsti:taughtBy dsti:FabienGandon ;
dsti:partOf ?programme ;
dsti:covers dsti:KnowledgeGraphs .
}Result
| programme |
|---|
| MSc in Data Science & AI |
| MSc in Data Engineering for AI |
| MSc in Data Analytics with AI |
| Executive MSc in Artificial Intelligence for Digital Transformation |
05 Pr Fabien Gandon: a research life built around the intelligence of the Web
Fabien Gandon is a Research Director at Inria, a computer scientist and an engineer in mathematics. His 2002 PhD pioneered the joint use of distributed artificial intelligence and Semantic Web technologies to manage diverse data sources and users over a Web architecture. At Carnegie Mellon University, he worked on AI methods for enforcing privacy preferences when querying and reasoning over personal data.
He founded the Wimmics research team in Sophia Antipolis and led it from 2012 to 2024, building a research programme that bridges formal semantics, social semantics and human–machine interaction. He has represented Inria at the W3C since 2012, holds a 3IA Côte d'Azur chair, and has chaired major international conferences across the Web and Semantic Web communities.
His work has moved repeatedly between theory, standards and deployment: DBpedia.fr and cultural Linked Data; DataLift for publishing and interlinking datasets; CovidOnTheWeb for biomedical literature; D2KAB for agronomy and biodiversity; WASABI for music knowledge; and HyperAgents for communities of people and artificial agents operating through Web architecture.
Artificial intelligence, knowledge graphs, knowledge representation, Linked Data, ontologies and distributed intelligence.
Inria representative at W3C, former Wimmics leader, conference chair and contributor to the evolution of the Web.
Co-President of the DSTI Scientific Advisory Board with Dr Christine Malot, and a DSTI professor since 2015.
Fabien is also a co-author of the third edition of Semantic Web for the Working Ontologist, a reference that carries the same philosophy as his teaching: formal ideas become valuable when they can be modelled, queried and used by working systems.
06 At DSTI: students build the semantic layers themselves
Semantic Web Technologies is taught across all DSTI Data MSc programmes and the Executive MSc in Artificial Intelligence for Digital Transformation. The course moves from the architecture of the Web to practical graph publication, validation, querying, ontology design and inference.
Web architecture and Linked Data
Understand the principles and standards that allow resources and datasets to be published, identified and connected globally.
RDF and Turtle
Model entities and relationships as graphs, use RDF's advanced features and express the results in practical syntaxes.
SHACL validation
Define structural schemata and verify that graph data satisfy explicit quality and integrity expectations.
SPARQL querying
Select, aggregate and transform graph data through an open standard query language designed for graph patterns.
Ontologies and inference
Use RDFS, OWL and SKOS to design vocabularies, formalise domain knowledge and derive additional statements.
Provenance and bridges
Connect graph data to other formats and models, while describing datasets, catalogues and their history with open vocabularies.
The learning rhythm combines preparatory work, self-evaluation, intensive courses and lab sessions, a substantial assignment and an examination. Students finish with more than terminology: they have published, validated, queried and reasoned over knowledge graphs.
07 DSTI's position: intelligence grows when learning and knowledge work together
Machine learning is exceptionally effective at extracting regularities from examples. Semantic technologies are exceptionally effective at representing identities, relationships, domain rules, provenance and shared definitions. Engineering judgement begins with understanding which capability the problem requires.
Learn from evidence
Use statistical and machine-learning models for patterns that are uncertain, too complex to state directly or expected to evolve with new observations.
Represent what is known
Use graphs, vocabularies, constraints and logic for knowledge that should remain explicit, interoperable, inspectable and reusable.
This hybrid position matters for enterprise data integration, scientific knowledge, regulatory compliance, cultural heritage, digital twins, intelligent agents and any system expected to remain understandable beyond one model version or one vendor platform.
08 Follow the research, standards and teaching trail
Fabien's own pages form a connected map of a career: research papers, standards work, projects, invited talks, outreach and teaching. The following starting points make that map easy to explore.