DigoryS's picture
Add graph JSONL converter and Neo4j loader (#2)
fb9a4be
|
Raw History Blame Contribute Delete
4.52 kB

Scripts

Load the Eedi Misconceptions Graph into a graph database in one command.

Script What it does Needs
csv_to_jsonl.py Converts the two CSVs into property-graph JSONL (nodes.jsonl, relationships.jsonl) Python 3.8+ only
load_eedi_graph.py Loads the JSONL into Neo4j (local or Aura); builds the JSONL first with --download pip install neo4j python-dotenv, Neo4j 5+

Quick start

pip install neo4j python-dotenv
python scripts/load_eedi_graph.py --download

The CSVs are read from data/ if you've cloned this repo. Otherwise they're downloaded from the v1.0 release (pick a revision with --tag). You can also download just the two scripts and run them on their own.

Connecting to Neo4j

Copy .env.example to .env and fill in your details. Credentials are read from environment variables, or from the nearest .env file in the script's folder or any folder above it (failing that, the folder you run it from, or above):

NEO4J_URI=bolt://localhost:7687      # or neo4j+s://xxxx.databases.neo4j.io for Aura
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=...
NEO4J_DATABASE_EEDI=eedi-v1          # optional; this is the default

To keep several instances in one .env, add a suffix, e.g. NEO4J_URI_LOCAL, NEO4J_PASSWORD_LOCAL, and choose with --target LOCAL. When more than one is configured, the script refuses to guess.

The database is created if it doesn't exist. This needs Neo4j Enterprise, which Neo4j Desktop includes. Aura (including the free tier) and Community edition have a single database, so pass --database neo4j to load into it. The graph fits comfortably within Aura Free's limits.

Useful options

Option Effect
--download (Re)build the JSONL from the CSVs before loading
--target NAME Use the NEO4J_*_NAME variables
--database NAME Load into this database (default eedi-v1)
--wipe Delete everything in the database first
--allow-nonempty Merge into a database that already has data
--no-derived Skip the derived count properties

By default the loader refuses to write to a database that already contains data.

Graph model

(:Subject)-[:SUBJECT_CONTAINS_TOPIC]->(:Topic)-[:TOPIC_CONTAINS_SUBTOPIC]->(:SubTopic)
(:SubTopic)-[:SUBTOPIC_CONTAINS_CONSTRUCT]->(:Construct)
(:Construct)-[:CONSTRUCT_HAS_MISCONCEPTION {percentage}]->(:Misconception)
Label Node id Properties
Subject eedi:subject:32 subjectSubjectId, name
Topic eedi:topic:144 topicSubjectId, name
SubTopic eedi:subtopic:202 subTopicSubjectId, name
Construct eedi:construct:27 constructId, name, misconceptionCount*
Misconception eedi:misconception:13708 misconceptionId, name, constructCount*, subTopicCount*, topicCount*, subjectCount*

* Computed by the loader, not present in the CSVs.

  • Field definitions: properties are the CSV columns in camelCase (ConstructId → constructId, ConstructName → name). See docs/data-dictionary.md for what each field means.
  • Node ids include the type. Construct ids overlap with Subject/Topic/SubTopic ids, so the type prefix keeps every id unique. In Neo4j the id is stored as uri, and every node also has the label EediNode.
  • percentage is PercentageMisconceptionExistsInThisConstruct: the share of a misconception's taggings that fall in that construct. It sums to 100 per misconception.
  • Full-text search on names: CALL db.index.fulltext.queryNodes('eedi_text', 'place value').

Version 1.0 loads 10,665 nodes and 21,838 relationships, and the loader checks these counts when it finishes.

Example queries

Misconceptions that appear in all four subjects:

MATCH (m:Misconception) WHERE m.subjectCount = 4
RETURN m.name, m.topicCount ORDER BY m.topicCount DESC

Where a misconception shows up:

MATCH (m:Misconception {misconceptionId: 13708})<-[r:CONSTRUCT_HAS_MISCONCEPTION]-(c:Construct)
      <-[:SUBTOPIC_CONTAINS_CONSTRUCT]-(st:SubTopic)
RETURN st.name, c.name, r.percentage ORDER BY r.percentage DESC

Licence

The data is CC BY 4.0. See the main README for attribution. The JSONL is a format conversion of the CSVs, so the same licence and attribution apply.

The scripts themselves are licensed under the MIT Licence (see LICENSE).