Knowledge Graphs

Most AI you have met in this track learns fuzzy patterns from data. Knowledge graphs are the other half of intelligence: structured, explicit knowledge — facts and how they connect — that a machine can store, query, and reason over precisely. They are how a system knows that Paris is the capital of France and France is in Europe, and can therefore answer "is Paris in Europe?" without ever being told directly.

Imagine storing what you know about the world. A list of sentences works for lookup but not for connecting ideas. A knowledge graph instead stores each fact as a tiny link between two things:

   [Paris] ──is capital of──► [France] ──is in──► [Europe]
      │                          │
   located in              has currency
      ▼                          ▼
  [Île-de-France]            [Euro]

Each arrow is a fact: Paris — is capital of — France. Three parts: two things and the relationship between them (often called a triple). String millions of these together and you get a map of knowledge a computer can walk across.

The power is in the connections. Even though nobody stored "Paris is in Europe," the machine can follow the path — Paris → France → Europe — and work it out. That is simple reasoning, and loose text cannot do it reliably.

This connects to your MSc-track theme: there are two kinds of "knowing." LLMs learn fuzzy patterns and are brilliant but unreliable on exact facts. Knowledge graphs store exact facts and relationships and are precise but must be built. The best systems use both.

A knowledge graph can answer questions that need chains of facts: "Which European capitals use the euro?" requires following capital → country → continent and country → currency, then combining. It can also enforce rules: if every person has exactly one birth date, the graph can flag a contradiction. This is the formal, rule-based side of AI — the same family as the Mars-rover knowledge-representation idea — and it is making a strong comeback because it shores up the weaknesses of LLMs.

These are not academic curiosities — they quietly run products you use constantly:

• Search panels — when a search engine shows a box of facts about a person or place (born, height, spouse, films), that comes from a knowledge graph, not a web page.
• Recommendations — "because you watched X" works by linking films, actors, genres, and your history in a graph and following the connections.
• Shopping — "frequently bought together" and product relationships sit in a graph.
• Fraud & security — linking people, accounts, devices, and transactions reveals suspicious patterns of connection a flat list would hide.

The most exciting use today is pairing graphs with language models to fix the LLM's biggest flaw: making up facts. Instead of trusting the model's fuzzy memory, you let it consult the graph for the precise, current truth — a structured cousin of the RAG idea from earlier in the track.

The pattern: the user asks a question, the system pulls the exact relevant facts from the knowledge graph, hands them to the LLM, and the model answers from those verified facts — in fluent language. You get the LLM's wonderful communication and the graph's reliability, with answers you can trace back to real, connected facts.