Neo4j GDS Certification: Workflows study guide

What the Workflows domain tests

This domain tests end-to-end GDS work: projecting graphs, managing the graph catalog, training models, and orchestrating machine-learning pipelines. You need to know the order of operations and which catalog stores what.

Core rules

The graph catalog holds named projected graphs. gds.graph.project creates one from labels and relationship types; Cypher projections support computed properties. gds.graph.list and gds.graph.drop manage lifecycle.
The model catalog stores trained models and pipeline objects, not data graphs. Use gds.model.list and gds.model.drop.
A pipeline chains steps: node property steps (for example FastRP embeddings), feature steps, relationship types, train/test splitting, feature selection, and a classifier.
Training produces a model that must exist before predict. Predictions require the same feature configuration used at training time.

Worked example

A link-prediction workflow:

CALL gds.graph.project('social', 'Person', 'KNOWS');
CALL gds.beta.pipeline.linkPrediction.create('lp');
CALL gds.beta.pipeline.linkPrediction.addNodeProperty('lp', 'fastRP', {embeddingDimension: 64, mutateProperty: 'embedding'});
CALL gds.beta.pipeline.linkPrediction.addFeature('lp', 'hadamard', {nodeProperties: ['embedding']});
CALL gds.beta.pipeline.linkPrediction.configureSplit('lp', {testFraction: 0.2});
CALL gds.beta.pipeline.linkPrediction.addLogisticRegression('lp');
CALL gds.beta.pipeline.linkPrediction.train('social', {pipeline: 'lp', modelName: 'lp-model'});

Reasoning: the graph must contain the relationship type being predicted; the embedding is mutated into the projected graph by the node property step; the feature step consumes that mutation; the split reserves test data; training writes a model into the model catalog.

Traps

Assuming the model catalog stores projected graphs, or the graph catalog stores models.
Running predict before training a model.
Forgetting that pipeline node-property steps mutate into the projected graph and must be paired with feature steps.
Dropping a projected graph while a pipeline still needs it for training.