Whole-Graph Projections in Neo4j Bloom: Visual Graph Analytics at Scale
Neo4jBloomGraph-ProjectionsGDSGraph-VisualizationGraph-AnalyticsData-ScienceVisual-Analytics
Loading…
Teaser: Neo4j Bloom now runs GDS algorithms — PageRank, Louvain, Betweenness Centrality, Shortest Path, FastRP — directly from the visual interface, projecting results as interactive graph layers. No Cypher, no Python, no context switching. This article covers the 18 available algorithms, the Bloom projection workflow, and a complete fraud ring detection example from exploration to export.
In July 2026, Neo4j launched Whole-Graph Projections in Bloom — a feature that lets analysts run Graph Data Science (GDS) algorithms directly from the visual query interface and explore the results as interactive graph projections.
Previously, running GDS algorithms required switching between Neo4j Browser (for Cypher), the GDS Python client (for algorithm configuration), and potentially a separate visualisation tool (for results). Whole-Graph Projections collapses this workflow into a single Bloom session: select nodes and relationships, choose an algorithm, configure parameters through a visual form, and immediately see the results rendered as a new graph projection.
For teams using Bloom as their primary graph exploration tool, this bridges the gap between visual ad-hoc querying and algorithmic graph analytics.
Bloom has long excelled at interactive graph exploration — searching nodes, expanding relationships, styling subgraphs. GDS, meanwhile, provides 50+ graph algorithms (centrality, community detection, pathfinding, node embedding) but has historically required programmatic access through Cypher CALL procedures or the Python client.
The workflow friction was significant:
Before Whole-Graph Projections:
1. Bloom: Explore → find interesting subgraph
2. Switch to Browser: Run CALL gds.graph.project(...)
3. Switch to Python/Browser: CALL gds.pageRank.stream(...)
4. Export results → re-import into Bloom
5. Style nodes by PageRank score → analyse
After Whole-Graph Projections:
1. Bloom: Explore → find interesting subgraph
2. Bloom: Right-click → "Run Graph Algorithm"
3. Bloom: Configure PageRank → "Project Results"
4. Analyse the projected graph in-place
Whole-Graph Projections eliminates steps 2–4, keeping the analyst inside Bloom for the entire workflow.
Whole-Graph Projections extends Bloom's rendering engine to support GDS algorithm results as first-class visual projections. The flow is:
Start with any Bloom query — a pattern match, a neighbourhood expansion, or the entire graph. Bloom's existing search and expansion tools work unchanged.
The core workflow follows a simple loop:
graph LR
A[Explore in Bloom] --> B[Select subgraph]
B --> C[Choose algorithm]
C --> D[Configure parameters]
D --> E[Run & Project]
E --> F[Interact with projection]
F -->|Refine| B
F -->|Export| G[Subgraph / CSV]
classDef action fill:#4C78A8,stroke:#2c4e6e,color:#fff
classDef result fill:#54A24B,stroke:#3a7a35,color:#fff
class A,B,C,D action
class E,F result
Right-click the canvas and choose Run Graph Algorithm. Bloom presents a dialog with available GDS algorithms:
┌──────────────────────────────────────────┐
│ Run Graph Algorithm │
│ │
│ Algorithm: ┌─────────────────────────┐ │
│ │ PageRank │ │
│ │ Betweenness Centrality │ │
│ │ Louvain (Community) │ │
│ │ Label Propagation │ │
│ │ Shortest Path │ │
│ │ Node Similarity │ │
│ │ FastRP │ │
│ │ GraphSAGE │ │
│ └─────────────────────────┘ │
│ │
│ Run on: ◎ Current selection │
│ ○ All visible nodes │
│ ○ Entire graph │
│ │
│ [Mutate] [Stream] [Project Results] │
└──────────────────────────────────────────┘
Each algorithm exposes its GDS configuration parameters through a visual form. For Louvain community detection:
┌──────────────────────────────────────────┐
│ Louvain Configuration │
│ │
│ Relationship weight: ○ None │
│ ○ Use property: │
│ ┌──────────────┐ │
│ │ transactionAmt│ │
│ └──────────────┘ │
│ Max levels: [3] │
│ Max iterations: [50] │
│ Tolerance: [0.0001] │
│ Include intermediate: ☐ │
│ │
│ ┌──────────────────────────────────────┐│
│ │ Run & Project ││
│ └──────────────────────────────────────┘│
└──────────────────────────────────────────┘
When the algorithm completes, Bloom creates a projected graph layer — a new visualisation that overlays the algorithm results onto the existing graph. Nodes are coloured and sized by algorithm scores:
The projection is interactive — you can expand nodes from the projected graph, run additional queries, or apply the results as a node property filter.
Whole-Graph Projections supports 18 GDS algorithms at launch, grouped into four categories:
| Algorithm | What It Measures | Bloom Visualisation |
|---|---|---|
| PageRank | Node importance via inbound links | Node size, colour ramp |
| Betweenness Centrality | Node bridge/connector value | Highlighted markers |
| Closeness Centrality | Node reachability | Radial gradient |
| Degree Centrality | Connection count | Node size |
| Algorithm | What It Finds | Bloom Visualisation |
|---|---|---|
| Louvain | Hierarchical communities | Colour-coded clusters |
| Label Propagation | Overlapping communities | Colour-coded, blended |
| Leiden | Improved Louvain | Colour-coded clusters |
| Weakly Connected Components | Disconnected subgraphs | Region highlighting |
| Triangle Count | Local clustering | Node badge |
| Algorithm | What It Computes | Bloom Visualisation |
|---|---|---|
| Shortest Path (A*) | Optimal path between nodes | Path highlighting |
| All Pairs Shortest Path | Distance matrix | Heat map overlay |
| Minimum Spanning Tree | Minimal connection tree | Tree layout |
| BFS/DFS | Traversal order | Step-by-step animation |
| Algorithm | What It Produces | Bloom Visualisation |
|---|---|---|
| FastRP | Dense node vectors | Similarity-based layout |
| GraphSAGE | Inductive embeddings | Subgraph similarity |
| Node2Vec | Random-walk embeddings | 2D projection |
Here's a complete workflow using Whole-Graph Projections to detect potential fraud rings:
1. Load the transaction graph — A Bloom search for MATCH (p:Person)-[:TRANSFERRED]->(a:Account) returns the core transaction network.
2. Run Louvain community detection — Right-click → Run Graph Algorithm → Louvain → Run & Project. Bloom colours nodes by community.
3. Identify suspicious clusters — One community shows 15 accounts all transferring to a single account. The cluster is visually distinct from the rest of the graph.
4. Run PageRank on the suspicious community — Select the cluster, right-click → PageRank. Several accounts have anomalously high PageRank scores relative to their transaction volume — a known fraud indicator.
5. Compute shortest paths — Select two suspected fraudulent accounts → Shortest Path → identify intermediary accounts that connect them.
6. Export the projection — The full projected graph (with algorithm scores as node properties) can be exported as a subgraph definition for downstream systems.
Whole-Graph Projections runs algorithms on the GDS execution engine, not in the browser. Only the projected results are sent to Bloom:
| Graph Size | Louvain Execution | Projection Transfer | Bloom Render |
|---|---|---|---|
| 10K nodes, 50K edges | ~200ms | ~50ms | ~100ms |
| 100K nodes, 500K edges | ~1.5s | ~200ms | ~400ms |
| 1M nodes, 5M edges | ~12s | ~1.5s | ~3s |
The "Run on: Entire graph" option supports up to 10 million nodes in the current preview, with larger graphs supported through sampling or subgraph selection.
| Limitation | Details |
|---|---|
| Node-aligned algorithms only | Relationship-level algorithms (Link Prediction, Common Neighbours) projected as node properties; edge visual projections in development. |
| Single algorithm per projection | Multi-algorithm pipelines require sequential projections. |
| Preview storage | Projected graphs are ephemeral — visible only in the current Bloom session. Use MATERIALIZE to persist. |
| Graph size limit | 10M nodes for "Entire graph" mode. Subgraph selection recommended for larger graphs. |
| Feature | Whole-Graph Projections | Neo4j Browser + GDS | Python GDS Client | Gephi |
|---|---|---|---|---|
| Visual algorithm config | ✅ Visual forms | ❌ Cypher only | ❌ Code only | ⚠️ GUI |
| In-place visualisation | ✅ Native Bloom | ❌ Separate export | ❌ Separate export | ✅ Native |
| Interactive expansion | ✅ Post-projection | ❌ | ❌ | ⚠️ Limited |
| Result styling | ✅ Automatic | ❌ Manual | ❌ Manual | ✅ Automatic |
| Multi-algorithm workflow | ⚠️ Sequential | ✅ Scriptable | ✅ Scriptable | ⚠️ Sequential |
| Export | ✅ Subgraph, CSV | ✅ CSV, JSON | ✅ All formats | ✅ All formats |
Whole-Graph Projections in Bloom represents a maturing of the graph analytics toolchain. By bringing GDS algorithms into the visual exploration environment, Neo4j closes the loop between "I see something interesting" and "let me analyse it algorithmically" — without requiring the analyst to leave Bloom or write a single line of Cypher or Python.
For graph data scientists, the feature reduces iteration time on algorithm selection and parameter tuning. For business analysts who work primarily in Bloom, it unlocks analytical capabilities that previously required developer intervention. The result is a tighter feedback loop between human pattern recognition and computational graph analysis.
Whole-Graph Projections is available on Neo4j AuraDB Enterprise and Neo4j 5.26+ as part of the Bloom 2026.07 release. Enable it from the Bloom settings panel under Experimental Features.