Architecture

AuraDS makes it easy to run graph algorithms on Neo4j by integrating two main components:

  • Neo4j Database, where graph data are loaded and stored, and Cypher queries and all database operations (for example user management, query termination, etc.) are executed;

  • Graph Data Science, a software component installed in the Neo4j Database, whose main purpose is to run graph algorithms on in-memory projections of Neo4j Database data.

Neo4j Graph Data Science concepts

The Neo4j Graph Data Science (GDS) library includes procedures to project and manage graphs, run algorithms, and train machine learning models.

Graph Catalog

The graph catalog is used to store and manage projected graphs via GDS procedures.

Algorithms

GDS contains many graph algorithms, invoked as Cypher procedures and run on projected graphs.

GDS algorithms are broken down into three tiers of maturity:

  • Alpha: experimental algorithms that may be changed or removed at any time. Algorithms in this tier are prefixed with gds.alpha.<algorithm>.

  • Beta: algorithms promoted from the Alpha tier to candidates for the Production tier. Algorithms in this tier are prefixed with gds.beta.<algorithm>.

  • Production: algorithms that have been rigorously tested for stability and scalability. Algorithms in this tier are prefixed with gds.<algorithm>.

Model Catalog

Some machine learning algorithms (for example Node Classification and GraphSage) need to use trained models in their computation. The model catalog is used to store and manage named trained models.

Pipeline Catalog

The pipeline catalog is used to manage machine learning pipelines. A pipeline groups together all the stages of a supported machine learning task (for example Node classification), from graph feature extraction to model training, in a single end-to-end workflow.

Graph data flow

Since GDS algorithms can only run in memory, the typical data flow involves:

  1. Reading the graph data from Neo4j Database

  2. Loading (projecting) the data into an in-memory graph

  3. Running an algorithm on a projected graph

  4. Writing the results back to Neo4j Database (if the algorithm runs in write mode)

architecture