Training the pipeline

This feature is in the beta tier. For more information on feature tiers, see API Tiers.

The train mode, gds.beta.pipeline.nodeClassification.train, is responsible for splitting data, feature extraction, model selection, training and storing a model for future use. Running this mode results in a classification model of type NodeClassification, which is then stored in the model catalog. The classification model can be applied to a possibly different graph which classifies nodes.

More precisely, the training proceeds as follows:

Apply the node property steps, added according to Adding node properties, on the graph. The graph filter on each step consists of contextNodeLabels + targetNodeLabels and contextRelationships + relationshipTypes.
Apply the targetNodeLabels filter to the graph.
Select node properties to be used as features, as specified in Adding features.
Split the input graph into two parts: the train graph and the test graph. This is described in Configuring the node splits. These graphs are internally managed and exist only for the duration of the training.
Split the nodes in the train graph using stratified k-fold cross-validation. The number of folds k can be configured as described in Configuring the node splits.
Each model candidate defined in the parameter space is trained on each train set and evaluated on the respective validation set for every fold. The evaluation uses the specified primary metric.
Choose the best performing model according to the highest average score for the primary metric.
Retrain the winning model on the entire train graph.
Evaluate the performance of the winning model on the whole train graph as well as the test graph.
Retrain the winning model on the entire original graph.
Register the winning model in the Model Catalog.

The above steps describe what the procedure does logically. The actual steps as well as their ordering in the implementation may differ.

A step can only use node properties that are already present in the input graph or produced by steps, which were added before.

Parallel executions of the same pipeline on the same graph is not supported.

Metrics

The Node Classification model in the Neo4j GDS library supports the following evaluation metrics:

Global metrics
- F1_WEIGHTED
- F1_MACRO
- ACCURACY
- OUT_OF_BAG_ERROR (only for RandomForest and only gives validation and test score)
Per-class metrics
- F1(class=<number>) or F1(class=*)
- PRECISION(class=<number>) or PRECISION(class=*)
- RECALL(class=<number>) or RECALL(class=*)
- ACCURACY(class=<number>) or ACCURACY(class=*)

The * is syntactic sugar for reporting the metric for each class in the graph. When using a per-class metric, the reported metrics contain keys like for example ACCURACY_class_1.

More than one metric can be specified during training but only the first specified — the primary one — is used for evaluation, the results of all are present in the train results. The primary metric may not be a * expansion due to the ambiguity of which of the expanded metrics should be the primary one.

The OUT_OF_BAG_ERROR is computed only for RandomForest models and is evaluated as the accuracy of majority voting, where for each example only the trees that did not use that example during training are considered. The proportion the train set used by each tree is controlled by the configuration parameter numberOfSamplesRatio. OUT_OF_BAG_ERROR is reported as a validation score when evaluated during the cross-validation phase. In the case when a random forest model wins, it is reported as a test score based on retraining the model on the entire train set.

Syntax

Run Node Classification in train mode on a named graph:

CALL gds.beta.pipeline.nodeClassification.train(
  graphName: String,
  configuration: Map
) YIELD
  trainMillis: Integer,
  modelInfo: Map,
  modelSelectionStats: Map,
  configuration: Map

Table 1. Parameters
Name	Type	Default	Optional	Description
graphName	String	`n/a`	no	The name of a graph stored in the catalog.
configuration	Map	`{}`	yes	Configuration for algorithm-specifics and/or graph filtering.

Table 2. Configuration
Name	Type	Default	Optional	Description
pipeline	String	`n/a`	no	The name of the pipeline to execute.
targetNodeLabels	List of String	`['*']`	yes	Filter the named graph using the given node labels to obtain nodes that are subject to training and evaluation.
relationshipTypes	List of String	`['*']`	yes	Filter the named graph using the given relationship types.
concurrency	Integer	`4`	yes	The number of concurrent threads used for running the algorithm.
targetProperty	String	`n/a`	no	The class of the node. Must be of type Integer.
metrics	List of String	`n/a`	no	Metrics used to evaluate the models.
randomSeed	Integer	`n/a`	yes	Seed for the random number generator used during training.
modelName	String	`n/a`	no	The name of the model to train, must not exist in the Model Catalog.
jobId	String	`Generated internally`	yes	An ID that can be provided to more easily track the training’s progress.
storeModelToDisk	Boolean	`false`	yes	Automatically store model to disk after training.

Table 3. Results
Name	Type	Description
trainMillis	Integer	Milliseconds used for training.
modelInfo	Map	Information about the training and the winning model.
modelSelectionStats	Map	Statistics about evaluated metrics for all model candidates.
configuration	Map	Configuration used for the train procedure.

The modelInfo can also be retrieved at a later time by using the Model List Procedure. The modelInfo return field has the following algorithm-specific subfields:

Table 4. Fields of modelSelectionStats
Name	Type	Description
bestParameters	Map	The model parameters which performed best on average on validation folds according to the primary metric.
modelCandidates	List	List of maps, where each map contains information about one model candidate. This information includes the candidates parameters, training statistics and validation statistics.
bestTrial	Integer	The trial that produced the best model. The first trial has number 1.

Table 5. Fields of modelInfo
Name	Type	Description
modelName	String	The name of the trained model.
modelType	String	The type of the trained model.
classes	List of Integer	Sorted list of class ids which are the distinct values of `targetProperty` over the entire graph.
bestParameters	Map	The model parameters which performed best on average on validation folds according to the primary metric.
metrics	Map	Map from metric description to evaluated metrics for the winning model over the subsets of the data, see below.
nodePropertySteps	List of Map	Algorithms that produce node properties within the pipeline.
featureProperties	List of String	Node properties selected as input features to the pipeline model.

The structure of modelInfo is:

{
    bestParameters: Map,                (1)
    nodePropertySteps: List of Map,
    featureProperties: List of String,
    classes: List of Integer,           (2)
    metrics: {                          (3)
        <METRIC_NAME>: {                (4)
            test: Float,                (5)
            outerTrain: Float,          (6)
            train: {                    (7)
                avg: Float,
                max: Float,
                min: Float,
            },
            validation: {               (8)
                avg: Float,
                max: Float,
                min: Float,
                params: Map
            }
        }
    }
}

1	The best scoring model candidate configuration.
2	Sorted list of class ids which are the distinct values of `targetProperty` over the entire graph.
3	The `metrics` map contains an entry for each metric description, and the corresponding results for that metric.
4	A metric name specified in the configuration of the procedure, e.g., `F1_MACRO` or `RECALL(class=4)`.
5	Numeric value for the evaluation of the winning model on the test set.
6	Numeric value for the evaluation of the winning model on the outer train set.
7	The `train` entry summarizes the metric results over the `train` set.
8	The `validation` entry summarizes the metric results over the `validation` set.

In (5)-(7), if the metric is OUT_OF_BAG_ERROR, these statistics are not reported. The OUT_OF_BAG_ERROR is only reported in (8) as validation metric and only if the model is RandomForest.

In addition to the data the procedure yields, there’s a fair amount of information about the training that’s being sent to the Neo4j database’s logs as the procedure progresses.

For example, how well each model candidates perform is logged with info log level and thus end up the neo4j.log file of the database.

Some information is only logged with debug log level, and thus end up in the debug.log file of the database. An example of this is training method specific metadata - such as per epoch loss for logistic regression - during model candidate training (in the model selection phase). Please note that this particular data is not yielded by the procedure call.

Example

All the examples below should be run in an empty database.

The examples use Cypher projections as the norm. Native projections will be deprecated in a future release.

In this section we will show examples of running a Node Classification training pipeline on a concrete graph. The intention is to illustrate what the results look like and to provide a guide in how to make use of the model in a real setting. We will do this on a small graph of a handful of nodes representing houses. This is an example of Multi-class classification, the class node property distinct values determine the number of classes, in this case three (0, 1 and 2). The example graph looks like this:

The following Cypher statement will create the example graph in the Neo4j database:

CREATE
  (gold:House {color: 'Gold', sizePerStory: [15.5, 23.6, 33.1], class: 0}),
  (red:House {color: 'Red', sizePerStory: [15.5, 23.6, 100.0], class: 0}),
  (blue:House {color: 'Blue', sizePerStory: [11.3, 35.1, 22.0], class: 0}),
  (green:House {color: 'Green', sizePerStory: [23.2, 55.1, 0.0], class: 1}),
  (gray:House {color: 'Gray', sizePerStory: [34.3, 24.0, 0.0],  class: 1}),
  (black:House {color: 'Black', sizePerStory: [71.66, 55.0, 0.0], class: 1}),
  (white:House {color: 'White', sizePerStory: [11.1, 111.0, 0.0], class: 1}),
  (teal:House {color: 'Teal', sizePerStory: [80.8, 0.0, 0.0], class: 2}),
  (beige:House {color: 'Beige', sizePerStory: [106.2, 0.0, 0.0], class: 2}),
  (magenta:House {color: 'Magenta', sizePerStory: [99.9, 0.0, 0.0], class: 2}),
  (purple:House {color: 'Purple', sizePerStory: [56.5, 0.0, 0.0], class: 2}),
  (pink:UnknownHouse {color: 'Pink', sizePerStory: [23.2, 55.1, 56.1]}),
  (tan:UnknownHouse {color: 'Tan', sizePerStory: [22.32, 102.0, 0.0]}),
  (yellow:UnknownHouse {color: 'Yellow', sizePerStory: [39.0, 0.0, 0.0]}),

  // richer context
  (schiele:Painter {name: 'Schiele'}),
  (picasso:Painter {name: 'Picasso'}),
  (kahlo:Painter {name: 'Kahlo'}),

  (schiele)-[:PAINTED]->(gold),
  (schiele)-[:PAINTED]->(red),
  (schiele)-[:PAINTED]->(blue),
  (picasso)-[:PAINTED]->(green),
  (picasso)-[:PAINTED]->(gray),
  (picasso)-[:PAINTED]->(black),
  (picasso)-[:PAINTED]->(white),
  (kahlo)-[:PAINTED]->(teal),
  (kahlo)-[:PAINTED]->(beige),
  (kahlo)-[:PAINTED]->(magenta),
  (kahlo)-[:PAINTED]->(purple),
  (schiele)-[:PAINTED]->(pink),
  (schiele)-[:PAINTED]->(tan),
  (kahlo)-[:PAINTED]->(yellow);

With the graph in Neo4j we can now project it into the graph catalog to prepare it for the pipeline execution. We do this using a Cypher projection targeting the House and UnknownHouse labels. We will also project the sizeOfStory property to use as a model feature, and the class property to use as a target feature.

The following statement will project a graph using a Cypher projection and store it in the graph catalog under the name 'myGraph'.

MATCH (house:House|UnknownHouse)
RETURN gds.graph.project(
  'myGraph',
  house,
  null,
  {
    sourceNodeLabels: labels(house),
    targetNodeLabels: [],
    sourceNodeProperties: house { .sizePerStory, .class },
    targetNodeProperties: {}
  }
)

Memory Estimation

First off, we will estimate the cost of running the algorithm using the estimate procedure. This can be done with any execution mode. We will use the train mode in this example. Estimating the algorithm is useful to understand the memory impact that running the algorithm on your graph will have. When you later actually run the algorithm in one of the execution modes the system will perform an estimation. If the estimation shows that there is a very high probability of the execution going over its memory limitations, the execution is prohibited. To read more about this, see Automatic estimation and execution blocking.

For more details on estimate in general, see Memory Estimation.

The following will estimate the memory requirements for running the algorithm in train mode:

CALL gds.beta.pipeline.nodeClassification.train.estimate('myGraph', {
  pipeline: 'pipe',
  targetNodeLabels: ['House'],
  modelName: 'nc-model',
  targetProperty: 'class',
  randomSeed: 2,
  metrics: [ 'ACCURACY' ]
})
YIELD requiredMemory

Table 6. Results
requiredMemory
"[1264 KiB ... 1337 KiB]"

If a node property step does not have an estimation implemented, the step will be ignored in the estimation.

Train

In the following examples we will demonstrate running the Node Classification training pipeline on this graph. We will train a model to predict the class in which a house belongs, based on its sizePerStory property.

The following will train a model using a pipeline:

CALL gds.beta.pipeline.nodeClassification.train('myGraph', {
  pipeline: 'pipe',
  targetNodeLabels: ['House'],
  modelName: 'nc-pipeline-model',
  targetProperty: 'class',
  randomSeed: 1337,
  metrics: ['ACCURACY', 'OUT_OF_BAG_ERROR']
}) YIELD modelInfo, modelSelectionStats
RETURN
  modelInfo.bestParameters AS winningModel,
  modelInfo.metrics.ACCURACY.train.avg AS avgTrainScore,
  modelInfo.metrics.ACCURACY.outerTrain AS outerTrainScore,
  modelInfo.metrics.ACCURACY.test AS testScore,
  [cand IN modelSelectionStats.modelCandidates | cand.metrics.ACCURACY.validation.avg] AS validationScores

Table 7. Results
winningModel	avgTrainScore	outerTrainScore	testScore	validationScores
{batchSize=100, classWeights=[], focusWeight=0.0, learningRate=0.001, maxEpochs=500, methodName="LogisticRegression", minEpochs=1, patience=1, penalty=5.881039654, tolerance=0.001}	1.0	1.0	1.0	[0.8, 0.0, 0.5, 0.9, 0.8]

Here we can observe that the model candidate with penalty 5.881 performed the best in the training phase, with an ACCURACY score of 1 over the train graph as well as on the test graph. This model is one that the auto-tuning found. This indicates that the model reacted very well to the train graph, and was able to generalize well to unseen data. Notice that this is just a toy example on a very small graph. In order to achieve a higher test score, we may need to use better features, a larger graph, or different model configuration.

Providing richer contexts to node property steps

In the above example we projected a House subgraph without relationships and used it for training and testing. Much information in the original graph is not used. We might want to utilize more node and relationship types to generate node properties (and link features) and investigate whether it improves node classification. We can do that by passing in contextNodeLabels and contextRelationshipTypes when adding a node property step.

The following statement will project a graph containing the information about houses and their painters using a Cypher projection and store it in the graph catalog under the name 'paintingGraph'.

MATCH (house:House)
OPTIONAL MATCH (painter:Painter)-[r:PAINTED]->(house:House)
RETURN gds.graph.project(
  'paintingGraph',
  painter,
  house,
  {
    sourceNodeLabels: ['Painter'],
    targetNodeLabels: ['House'],
    sourceNodeProperties: {},
    targetNodeProperties: house { .class },
    relationshipType: 'PAINTED'
  },
  { undirectedRelationshipTypes: ['PAINTED'] }
)

We still train a model to predict the class of each house, but use Painter and PAINTED as context in addition to House to generate features that leverage the full graph structure. After the feature generation however, it is only the House nodes that are considered as training and evaluation instances, so only the House nodes need to have the target property class.

First, we create a new pipeline.

CALL gds.beta.pipeline.nodeClassification.create('pipe-with-context')

Second, we add a node property step (in this case, a node embedding) with Painter as contextNodeLabels.

CALL gds.beta.pipeline.nodeClassification.addNodeProperty('pipe-with-context', 'fastRP', {
embeddingDimension: 64,
iterationWeights: [0, 1],
mutateProperty:'embedding',
contextNodeLabels: ['Painter']
})

We add our embedding as a feature for the model:

CALL gds.beta.pipeline.nodeClassification.selectFeatures('pipe-with-context', ['embedding'])

And we complete the pipeline setup by adding a logistic regression model candidate:

CALL gds.beta.pipeline.nodeClassification.addLogisticRegression('pipe-with-context')

We are now ready to invoke the training of the newly created pipeline.

The following will train a model using the context-configured pipeline:

CALL gds.beta.pipeline.nodeClassification.train('paintingGraph', {
  pipeline: 'pipe-with-context',
  targetNodeLabels: ['House'],
  modelName: 'nc-pipeline-model-contextual',
  targetProperty: 'class',
  randomSeed: 1337,
  metrics: ['ACCURACY']
}) YIELD modelInfo, modelSelectionStats
RETURN
  modelInfo.bestParameters AS winningModel,
  modelInfo.metrics.ACCURACY.train.avg AS avgTrainScore,
  modelInfo.metrics.ACCURACY.outerTrain AS outerTrainScore,
  modelInfo.metrics.ACCURACY.test AS testScore,
  [cand IN modelSelectionStats.modelCandidates | cand.metrics.ACCURACY.validation.avg] AS validationScores

Table 8. Results
winningModel	avgTrainScore	outerTrainScore	testScore	validationScores
{batchSize=100, classWeights=[], focusWeight=0.0, learningRate=0.001, maxEpochs=100, methodName="LogisticRegression", minEpochs=1, patience=1, penalty=0.0, tolerance=0.001}	1.0	1.0	1.0	[1.0]

As we can see, the results indicate that the painter information is sufficient to perfectly classify the houses. The change is due to the embeddings taking into account more contextual information. While this is a toy example, additional context can sometimes provide valuable information to pipeline steps, resulting in better performance.