TopolearnCentrality, communities and similarity power recommendations and fraud detection, and plain Cypher cannot compute them efficiently.
The Graph Data Science (GDS) library is a Neo4j plugin (included in Aura Graph Analytics and Desktop) providing parallel implementations of more than 60 algorithms. Its workflow differs from normal Cypher:
- project a graph into memory:
gds.graph.project('name', nodeLabels, relTypes, config)builds a compact in-memory copy of just the nodes and relationships you need, optionally with properties as weights - run an algorithm in one of four modes:
stream(return results as rows),stats(summary only),mutate(write results into the in-memory graph for use by other algorithms),write(persist as properties back into the database) - use
estimatefirst on large graphs to check memory, andgds.graph.dropafterwards
Popular algorithm families:
- centrality: PageRank, betweenness, degree, to find influential or bridging nodes
- community detection: Louvain, Leiden, label propagation, weakly connected components for clusters and fraud rings
- similarity: node similarity and k-nearest neighbours for "people like you" recommendations
- path finding: Dijkstra, A* and Yen's k-shortest for weighted routes
- embeddings (FastRP, GraphSAGE) and link prediction pipelines for machine learning
Results are node ids; gds.util.asNode(nodeId) maps them back to database nodes. GDS has a Python client (graphdatascience) that wraps the same procedures.
Example
// 1. project Person nodes and FOLLOWS relationships into memory
CALL gds.graph.project('social', 'Person', 'FOLLOWS');
// 2. stream PageRank scores: most influential people first
CALL gds.pageRank.stream('social')
YIELD nodeId, score
RETURN gds.util.asNode(nodeId).name AS name, round(score, 3) AS score
ORDER BY score DESC LIMIT 5;
// 3. find communities and store them on the nodes
CALL gds.louvain.write('social', {writeProperty: 'community'})
YIELD communityCount, modularity;
// 4. free the memory
CALL gds.graph.drop('social');Remember
- GDS projects a graph into memory, then runs parallel algorithms on it
- Modes: stream, stats, mutate, write; estimate memory first
- Centrality, communities, similarity, weighted paths and embeddings
- Map node ids back with gds.util.asNode
Watch out for
- Forgetting to drop projected graphs, leaving gigabytes of heap occupied.