neo4j-gds-skill

When to Use

• Running GDS algorithms against embedded GDS plugin through Python client (graphdatascience)
• Running GDS algorithms through CALL gds.* Cypher procedures
• Aura Pro, self-managed Neo4j, local Neo4j, or offline DBMS with GDS plugin installed
• Projecting named in-memory graphs, running centrality/community/similarity/path/embedding algorithms
• Chaining algorithms via mutate mode; building FastRP → KNN pipelines
• Writing node embeddings for Neo4j vector indexes / structural similarity search
• Memory estimation before large graph operations

When NOT to Use

• Aura Graph Analytics Sessions / AGA / GdsSessions / AuraGraphDataScience → neo4j-aura-graph-analytics-skill
• AuraDB Cypher API with { memory: ... } or { sessionId: ... } → neo4j-aura-graph-analytics-skill
• Cypher query authoring → neo4j-cypher-skill
• Driver/connection setup → neo4j-driver-python-skill
• GraphRAG retrieval → neo4j-graphrag-skill
• Creating/querying vector indexes over written embeddings → neo4j-vector-index-skill

Pre-flight

Use only with embedded GDS plugin.

from graphdatascience import GraphDataScience

gds = GraphDataScience("neo4j+s://xxx.databases.neo4j.io", auth=("neo4j", "pw"))   # AuraDS; aura_ds auto-derived
gds = GraphDataScience("bolt://localhost:7687", auth=("neo4j", "password"))
print(gds.server_version())

RETURN gds.version() AS gds_version

GDS plugin unavailable: client raises GdsNotFound at construction; Cypher raises Unknown function 'gds.version'. AuraDB serverless analytics → neo4j-aura-graph-analytics-skill. Self-managed/local → install or enable GDS plugin.

pip install "graphdatascience>=2.1"          # 2.1 required for GDS 2026.09
pip install "graphdatascience[rust-ext]"     # optional: faster serialization

Compatibility: graphdatascience 2.1 / 2.2 — GDS >= 2.13 and < 2.28 / < 2026.10; 2.0 — GDS < 2026.9. All: Python >= 3.10 and < 3.15, Neo4j Python driver >= 5.26 and < 7.0, pandas 2–3, pyarrow 21–25. GDS server < 2.13 → DeprecationWarning at construction; pin graphdatascience<2 (client 1.22) there.

Client 1.x fallback (GDS server < 2.13)

Migration guide: Neo4j GDS Python client 2.0 migration

GDS plugin releases track the server: 2026.09.0 requires Neo4j 2026.09 — check the GDS compatibility table before upgrading either side.

GDS plugin 2026.07.0 removed CALL gds.userLog() — read hints and warnings from driver result summary notifications or the Neo4j debug log; track task progress with CALL gds.listProgress().

2.0 client rules:

• Plain endpoints, no v2 prefix — untyped 1.x endpoints and gds.v2.* are gone
• snake_case parameters: page_rank, fast_rp, mutate_property, write_property
• Typed result attributes: result.write_millis, not result["writeMillis"] (stream still returns DataFrame)
• Server version via gds.server_version() — no gds.version() client method (Cypher RETURN gds.version() still valid)
• Procedure aliases: gds.betweenness ≡ gds.betweenness_centrality; also gds.closeness, gds.degree, gds.eigenvector, gds.harmonic, gds.kcore
• Pipelines: gds.pipeline.node_classification / link_prediction / node_regression — the only API in 2.0
• gds.run_cypher(query, auto_commit=True) [2.1] for CALL { … } IN TRANSACTIONS — 2.0 default retryable transaction rejects it; on 2.0 use the Neo4j driver directly
• gds.db_driver() [2.1] → underlying neo4j.Driver for custom sessions/transactions; closed by gds.close() only if client created it
• mode="READ" / "WRITE" strings accepted wherever QueryMode is [2.1]
• No async/job-handle API on the plugin surface — compute(), *_async, gds.jobs are AGA Sessions only

Graph Catalog Operations

Native Projection

CALL gds.graph.project(
  'myGraph',
  ['Person', 'City'],
  { KNOWS: { orientation: 'UNDIRECTED' }, LIVES_IN: {} }
)
YIELD graphName, nodeCount, relationshipCount

G, result = gds.graph.project.native("myGraph", "Person", "KNOWS")
print(result.node_count, result.relationship_count)

G, result = gds.graph.project.native(
    "myGraph",
    {"Person": {"properties": ["age", "score"]}, "City": {}},
    {"KNOWS": {"orientation": "UNDIRECTED"}, "LIVES_IN": {"properties": ["since"]}},
    overwrite=True,   # drop same-named graph first
)

Native projection: plugin/simple Python-client workflow only. AGA Sessions → neo4j-aura-graph-analytics-skill. 1.x fallback: gds.v2.graph.project(...).

Cypher Projection (use for new Cypher workflows, filters, transforms)

G, result = gds.graph.project.cypher(
    """
    MATCH (source:Person)-[r:KNOWS]->(target:Person)
    WHERE source.active = true
    RETURN gds.graph.project($graph_name, source, target,
        { sourceNodeProperties: source { .score }, relationshipType: 'KNOWS' })
    """,
    graph_name="activeGraph",
)

gds.graph.project.cypher(query) takes no database= — set GraphDataScience(..., database=...) at construction or call gds.set_database(...) before projecting. Query must end with exactly one RETURN gds.graph.project(...). If validation fails: use gds.run_cypher(...), then gds.graph.get("graphName"). 1.x fallback: gds.graph.cypher.project(query, database=...).

AGA Sessions → neo4j-aura-graph-analytics-skill; never use plugin Cypher projection.

Undirected Projection

Native projection: set orientation: 'UNDIRECTED' per relationship type. Plugin Cypher projection: set undirectedRelationshipTypes: ['*'] in fifth gds.graph.project(...) config argument.

Leiden is defined for directed and undirected graphs. Project undirected relationships when community structure is naturally symmetric.

Inspect and Drop

G.node_count()              # 12_043
G.relationship_count()      # 87_211
G.node_properties()         # projected + mutated properties by label
G.relationship_properties() # projected + mutated properties by type
G.size_in_bytes()
gds.graph.drop(G)           # frees JVM heap; returns list[GraphInfo]

G = gds.graph.get("myGraph")       # re-attach to existing projection

gds.graph.list()

Memory Estimation — run before large projections and algorithms

CALL gds.graph.project.estimate(['Person'], 'KNOWS')
YIELD requiredMemory, bytesMin, bytesMax, nodeCount, relationshipCount

est = gds.graph.project.estimate(node_projection=["Person"], relationship_projection=["KNOWS"])
print(est.required_memory)

G, project_result = gds.graph.project.native("myGraph", "Person", "KNOWS")
print(project_result.node_count)

# Algorithm estimation:
est = gds.page_rank.estimate(G, damping_factor=0.85)
print(est.required_memory)

Execution Modes

Pattern: stream to verify → mutate to chain → write to persist.

mutate_property must not exist in the in-memory graph. Relationship algorithms such as KNN also require mutate_relationship_type. After write, re-project to use written properties in subsequent GDS calls (in-memory graph does not see DB writes).

gds.util.asNode() — Enrich Stream Results

stream mode yields nodeId (internal GDS integer). gds.util.asNode(nodeId) translates it back to the DB node so you can access properties.

// Single property
CALL gds.pageRank.stream('myGraph', {})
YIELD nodeId, score
RETURN gds.util.asNode(nodeId).name AS name, score
ORDER BY score DESC LIMIT 10

// Multiple properties — convert once with WITH
CALL gds.pageRank.stream('myGraph', {})
YIELD nodeId, score
WITH gds.util.asNode(nodeId) AS node, score
RETURN node.name AS name, node.born AS born, score
ORDER BY score DESC LIMIT 10

Not needed for write, mutate, or stats modes — those don't return per-node data.

Core Algorithms

PageRank (centrality)

CALL gds.pageRank.stream('myGraph', { dampingFactor: 0.85, maxIterations: 20 })
YIELD nodeId, score
RETURN gds.util.asNode(nodeId).name AS name, score ORDER BY score DESC LIMIT 10
// score: relative influence — not absolute. Compare within same run only.
// didConverge: true means score stabilized; if false, increase maxIterations.

CALL gds.pageRank.write('myGraph', { writeProperty: 'pagerank', dampingFactor: 0.85 })
YIELD nodePropertiesWritten, ranIterations, didConverge

pr_df = gds.page_rank.stream(G, damping_factor=0.85)
mutate_result = gds.page_rank.mutate(G, mutate_property="pagerank", damping_factor=0.85)
write_result = gds.page_rank.write(G, write_property="pagerank", damping_factor=0.85)
print(write_result.write_millis)

Louvain (community detection)

CALL gds.louvain.stream('myGraph', { relationshipWeightProperty: 'weight' })
YIELD nodeId, communityId

CALL gds.louvain.write('myGraph', { writeProperty: 'community' })
YIELD communityCount, modularity

louvain_df = gds.louvain.stream(G)
write_result = gds.louvain.write(G, write_property="community")
print(write_result.community_count)

Leiden is a refinement of Louvain avoiding poorly connected communities — use when community quality > raw speed. modularity in stats result: range -0.5 to 1.0. [field] Values > 0.3 often indicate meaningful community structure; > 0.7 is strong. Leiden is defined for directed and undirected graphs. Project undirected relationships when community structure is naturally symmetric.

WCC — Weakly Connected Components

Run WCC first to understand graph structure; partition disconnected graphs before expensive algorithms.

CALL gds.wcc.stream('myGraph', { minComponentSize: 10 })
YIELD nodeId, componentId

CALL gds.wcc.write('myGraph', { writeProperty: 'componentId' })
YIELD nodePropertiesWritten, componentCount

wcc_df = gds.wcc.stream(G)
write_result = gds.wcc.write(G, write_property="componentId")
print(write_result.node_properties_written)

Betweenness Centrality

gds.betweenness.stream(G)          # alias of betweenness_centrality; identifies bottleneck/bridge nodes
gds.betweenness.write(G, write_property="betweenness")

Node Similarity

Jaccard similarity from common neighbors — no node properties required.

gds.node_similarity.stream(G, similarity_cutoff=0.1, top_k=10)
gds.node_similarity.write(G, write_relationship_type="SIMILAR", write_property="score",
                          similarity_cutoff=0.1, top_k=10)

FastRP (node embeddings)

Fast, scalable, production ML pipelines. Set randomSeed for reproducibility.

CALL gds.fastRP.mutate('myGraph', {
  embeddingDimension: 256,
  iterationWeights: [0.0, 1.0, 1.0],
  featureProperties: ['score'],
  propertyRatio: 0.5,
  normalizationStrength: -0.5,
  randomSeed: 42,
  mutateProperty: 'embedding'
})
YIELD nodePropertiesWritten

gds.fast_rp.mutate(G, embedding_dimension=256, iteration_weights=[0.0, 1.0, 1.0],
                   random_seed=42, mutate_property="embedding")
write_result = gds.fast_rp.write(G, embedding_dimension=256, write_property="embedding",
                                 random_seed=42)
print(write_result.write_millis)

For ANN search over structural embeddings, after write, create a Neo4j vector index over the written property. Use neo4j-vector-index-skill.

KNN — K-Nearest Neighbors

Finds k most similar nodes per node based on node properties (typically embeddings).

CALL gds.knn.stream('myGraph', {
  nodeProperties: ['embedding'], topK: 10,
  sampleRate: 0.5, similarityCutoff: 0.7
})
YIELD node1, node2, similarity

CALL gds.knn.write('myGraph', {
  nodeProperties: ['embedding'], topK: 10,
  writeRelationshipType: 'SIMILAR', writeProperty: 'score'
})
YIELD relationshipsWritten

knn_df = gds.knn.stream(G, node_properties=["embedding"], top_k=10)
gds.knn.write(G, node_properties=["embedding"], top_k=10,
              write_relationship_type="SIMILAR", write_property="score")

FastRP → KNN Pipeline (recommendation)

# 1. Project
G, _ = gds.graph.project.native("myGraph", "Product",
    {"BOUGHT_TOGETHER": {"orientation": "UNDIRECTED"}})

# 2. Estimate memory
print(gds.fast_rp.estimate(G, embedding_dimension=128).required_memory)

# 3. Embed
gds.fast_rp.mutate(G, embedding_dimension=128, random_seed=42, mutate_property="emb")

# 4. Similarity
gds.knn.write(G, node_properties=["emb"], top_k=10,
              write_relationship_type="SIMILAR", write_property="score")

# 5. Cleanup
gds.graph.drop(G)

Algorithm Selection

Full algorithm catalog → references/algorithms.md

Common Errors

Full Workflow

1. Create gds with GraphDataScience(...).
2. Verify plugin: gds.server_version() or RETURN gds.version().
3. Estimate memory: gds.graph.project.estimate(...) and algorithm .estimate(...).
4. Project named graph with gds.graph.project.native(...) or gds.graph.project.cypher(query).
5. Run gds.*.stream first; switch to mutate; use write only when satisfied.
6. Drop graph with gds.graph.drop(G).
7. GDS server < 2.13 → client 1.22 via the mapping table above.

Built-in test datasets: gds.graph.datasets.load_cora(), gds.graph.datasets.load_karate_club(), gds.graph.datasets.load_imdb()

MCP Tool Mapping

Before any write-cypher: show exact Cypher, expected nodes/relationships affected, and ask for confirmation. For algorithm write mode, estimate or run stats first when available.

References

• references/algorithms.md — full algorithm catalog: all procedures, parameters, tiers, Cypher + Python examples
• references/graph-projection.md — projection deep-dive: filtering, heterogeneous graphs, relationship orientation, property types
• GDS Manual
• Python Client Docs

Checklist

• Embedded GDS plugin confirmed with gds.server_version() or RETURN gds.version()
• Graph/algorithm memory estimated before large work
• Python examples use 2.0 endpoints (no v2 prefix), snake_case params, typed result attributes
• Client 1.x used only with GDS server < 2.13, via the mapping table
• Projection uses native or plugin Cypher projection; no gds.graph.project.remote(...)
• Named graph dropped after use (gds.graph.drop(G); 1.x: gds.v2.graph.drop(G))
• Execution mode chosen: stream (inspect) → mutate (chain) → write (persist)
• write_property/mutate_property checked for collision with existing properties
• overwrite=True when re-projecting an existing graph name
• randomSeed set for reproducible embeddings
• WCC run first on graphs that may be disconnected