GraphSAGE node embedding training

GraphSAGE can be used as an unsupervised algorithm to generate embeddings for nodes in a graph. This page provides instructions for how to use the GraphSAGE node embedding training endpoint.

Syntax

This section covers the syntax used to execute the GraphSAGE node embedding training algorithm.

Run GraphSAGE node embedding training.
CALL graph.gs_unsup_train(
  'CPU_X64_XS',                    (1)
  {
    ['defaultTablePrefix': '...',] (2)
    'project': {...},              (3)
    'compute': {...},              (4)
  }
);
1 Compute pool selector.
2 Optional prefix for table references.
3 Project config.
4 Compute config.
Table 1. Parameters
Name Type Default Optional Description

computePoolSelector

String

n/a

no

The selector for the compute pool on which to run the GraphSAGE node embedding training job.

configuration

Map

{}

no

Configuration for graph project, algorithm compute and result write back.

For this algorithm we strongly recommend using a GPU compute pool, unless the dataset is very small and the model shallow.

The configuration map consists of the following three entries.

For more details on below Project configuration, refer to the Project documentation.
Table 2. Project configuration
Name Type

nodeTables

List of node tables.

relationshipTables

Map of relationship types to relationship tables.

Please note that in order for GraphSAGE to properly propagate updates of node embeddings, each type of node must be the target of at least one relationship type. The orientation parameter can be useful to add reverse direction relationships for types of nodes that are only the source of relationships (using the "REVERSE" or "UNDIRECTED" orientations).

Table 3. Compute configuration
Name Type Default Optional Description

numWalks

Integer

10

yes

The number of random walks to perform for each node in the graph

walkDepth

Integer

3

yes

The number of steps in every random walk

negSamplingRatio

Float

1.0

yes

The ratio of negative to positive samples to sample for training

modelname

String

n/a

no

The name of the model to train (must be unique)

numEpochs

Integer

n/a

no

The number of epochs to train the model

numSamples

List of Integer

n/a

no

The number of neighbors to sample for each layer. Note that this also determines the number of layers

hiddenChannels

Integer

256

yes

The node embedding dimension of the model layers' outputs

activation

String

"relu"

yes

The activation function to use. Valid values are "relu" and "sigmoid"

aggregator

String

"mean"

yes

The neighborhood embedding aggregator to use. Valid values are "mean" and "max"

learningRate

Float

0.001

yes

The learning rate for the optimizer

dropout

Float

0.1

yes

The dropout probability for each layer. Must be a value >= 0.0 and < 1.0

layerNormalization

Boolean

true

yes

Whether to apply layer normalization between the model layers

epochsPerCheckpoint

Integer

max(numEpochs / 10, 1)

yes

The number of epochs between saving model checkpoints

randomSeed

Integer

A random integer

yes

A number used to seed all randomness of the computation

batchSize

Integer

Automatically inferred

yes

The number of target nodes to train on in each batch. If not provided, the algorithm will automatically infer the maximally allowed batch size within the constraints of available memory

lossReduction

String

Automatically inferred

yes

The reduction method to apply to the loss. Valid values are "mean" and "sum". If not provided, the reduction method will "mean" if explicit batchSize is provided and otherwise "sum"

Example

For our example we will use an IMDB dataset with actors, directors, movies, and genres. These all have keywords associated with them, which we will use as features for the nodes. They are connected by relationships where actors act in movies and directors direct movies.

We have a database called imdb that contains the tables:

  • actor with columns nodeid and plot_keywords

  • movie with columns nodeid and plot_keywords

  • director with columns nodeid and plot_keywords

  • acted_in with columns sourcenodeid and targetnodeid that represent actor and movie node IDs

  • directed_in with columns sourcenodeid and targetnodeid that represent director and movie node IDs

The plot_keywords columns contain keywords associated with the nodes, encoded as vectors of floats.

You can upload this dataset to your snowflake account by following the instructions at github: neo4j-product-examples/snowflake-graph-analytics.

The training query

In the following query we train a GraphSAGE model for node embeddings on the dataset [1].

We train for 10 epochs, with two hidden layers.

To run the query, there is a required setup of grants for the application, your consumer role and your environment. Please see the Getting started page for more on this.

We also assume that the application name is the default Neo4j_Graph_Analytics. If you chose a different app name during installation, please replace it with that.

CALL Neo4j_Graph_Analytics.graph.gs_unsup_train('GPU_NV_S', {
    'defaultTablePrefix': 'imdb.gml',
    'project': {
        'nodeTables': ['actor', 'director', 'movie'],
        'relationshipTables': {
            'acted_in': {
                'sourceTable': 'actor',
                'targetTable': 'movie',
                'orientation': 'UNDIRECTED'
            },
            'directed_in': {
                'sourceTable': 'director',
                'targetTable': 'movie',
                'orientation': 'UNDIRECTED'
            }
        }
    },
    'compute': {
        'modelname': 'unsup-imdb',
        'numEpochs': 10,
        'numSamples': [20, 20]
    }
});

The above query should produce a row with empty job result.

JOB_ID

JOB_START

JOB_END

JOB_RESULT

job_c047364f8c3c4dc19f1e06fc3711483f

2025-04-29 12:39:08.215

2025-04-29 12:42:12.820

{}


1. We do not want to use the property genre of Movies when computing the embeddings because not all movies have genres, and moreover, using genre would make the embeddings inappropriate to use for predicting movie genres. A way to remove the genre property is to create a snowflake view of the movie table and select only nodeid and plot_keywords columns. Also, remember to grant the application SELECT privilege on the updated movie view. For simplicity, we keep the name movie and assume it does not have a genre column.