Ch 3. Triples and Graph Data Models (RDF vs LPG)¶
What you'll learn
- The minimal unit of knowledge — the triple (subject–predicate–object)
- Two graph data models — RDF (triple store) and LPG (labeled property graph)
- How the two hold the same fact differently, and when to pick which
- Why this course centers on LPG (Neo4j)
Assumes
Ch 2. You need the ontology (schema) vs. knowledge graph (instance) distinction in your head.
1. Concept — all knowledge breaks down into triples¶
The atom of a knowledge graph is the triple. Write a single fact as subject — predicate — object.
"Sujin Lee is employed at Apple" is one triple. Several sentences also break into a bundle of triples.
(Aspirin) —[treats]→ (headache)
(Aspirin) —[ingredient]→ (acetylsalicylic acid)
(headache) —[location]→ (head)
String these triples together and you get a graph. Subjects and objects are nodes, predicates are edges. A knowledge graph is, in the end, a giant graph where countless triples connect through shared nodes.
2. Why you need this — representation split into two camps¶
Historically, two ways of storing and querying the same triples diverged. Both are "node–edge–node," but they emphasize different things.
RDF (Resource Description Framework)¶
The semantic-web standard. It sees everything as pure triples. Nodes and predicates carry global identifiers (URIs), and to attach a property to a relationship itself you have to mint more triples (reification). You query with SPARQL and reason with OWL.
Strengths: standardization, interoperability, global identifiers. Strong in open data, life sciences, and knowledge integration.
LPG (Labeled Property Graph)¶
The industry standard, exemplified by Neo4j. It attaches properties (key-value) directly to both nodes and relationships. "Sujin Lee employed at Apple since 2019" — pinning a timestamp onto the relationship is natural. You query with Cypher.
Strengths: intuitive modeling, relationship properties, fast traversal. Popular in applications and real-time services.
3. Where it's used — same fact, different container¶
Comparing how the two hold "Sujin Lee has been employed at Apple since 2019" makes the difference sharp.
| RDF | LPG | |
|---|---|---|
| Property on a relation (since=2019) | Can't attach directly → needs reification | Just {since: 2019} on the relation |
| Identifiers | URIs (global) | Internal IDs (local) |
| Query language | SPARQL | Cypher |
| Reasoning (OWL) | Strong | Weak (at the app level) |
| Main use | Open / academic / data integration | Apps / real-time / recommendations |
If regular reasoning and global interoperability are central, RDF; if relationships carry many properties and intuitive modeling and fast traversal matter, LPG is more comfortable.
4. This course's choice — LPG/Neo4j¶
This course centers on LPG (Neo4j) for three reasons. Attaching properties to relationships fits GraphRAG's real-world patterns (timestamps, weights, source attribution) well; Cypher has a low barrier to entry; and the GraphRAG open-source ecosystem mostly clusters on the property-graph side. We mention RDF/SPARQL in context when relevant.
Neither one is 'correct'
The concept — representing knowledge as triples — is identical. The data model is a tooling choice, not a creed. The graph thinking you learn here carries straight over to RDF.
5. Common failure points¶
Unsure whether a relation should be a node or a property. When "employment" accrues a start date, a title, and a department, it's better to promote it from a relationship to its own node (e.g., Employment). A common modeling decision in LPG.
Splitting triples too finely. Turn every adjective into a triple and the graph explodes. The balance is to make only "the relationships a question needs to follow" into edges, and leave the rest as properties.
6. Exercises & next¶
Check your understanding¶
- Break 3 sentences from your domain into triples. What becomes a node, what becomes a relation, what becomes a property?
- If "employment" carries a start date, end date, and title, how would you express it in RDF vs. LPG?
Next¶
On to Part 2 — the step before triples: ontology modeling, where you design which concepts and relations to have → Part 2.
Sources¶
- W3C. RDF 1.1 Primer.
- Robinson, Webber & Eifrem (2015). Graph Databases (2nd ed.), O'Reilly
- Angles et al. (2017). Foundations of Modern Query Languages for Graph Databases. ACM Computing Surveys