GDS 파이프라인 및 머신 러닝

Welcome to GDS Pipelines! In the Neo4j Graph Data Science (GDS) library, a pipeline is a structured workflow for common graph data science tasks.

Think of it as a blueprint that defines a sequence of steps, from feature engineering using graph algorithms to training and deploying machine learning models.

GDS Pipelines offer several key benefits:

• Reproducibility: Define your entire workflow once and reuse it.
• Automation: Streamline complex tasks involving multiple graph algorithms and ML steps.
• Operationalization: Easily deploy graph-based machine learning models for continuous prediction.

They help bridge the gap between experimentation and production.

One common pipeline type is the Node Property Prediction Pipeline. This helps predict a specific property on nodes based on other node features and graph structure.

Let's create an empty pipeline named 'myChurnPrediction':

Before using a pipeline, you need to project your graph into GDS memory. This makes nodes and relationships accessible for algorithms and features.

Here, we project a simple 'User' graph with 'KNOWS' relationships:

Pipelines allow you to specify which existing node properties should be used as features for your machine learning model. These are direct attributes of your nodes.

Let's add 'age' and 'income' properties as features to our pipeline:

The real power of GDS pipelines is integrating graph algorithm results as features. These capture structural insights that simple node properties can't.

We can add PageRank scores as a feature to our pipeline:

After defining your features, you specify the machine learning model that will perform the prediction. GDS supports various models like Logistic Regression, Random Forest, and GNNs.

Let's add a Logistic Regression model to predict 'isChurned':

With features and a model defined, you can now train the pipeline. This executes all feature generation steps, then trains the specified ML model on the generated features.

We'll train our pipeline on 'mySocialGraph' and name the resulting model 'churnPredictionModel':

Once trained, your model (which is part of the pipeline) can be used to generate predictions on your graph data. You can either stream results or mutate the graph with new properties.

Let's stream predictions for our 'churnPredictionModel':

You can list all active pipelines and their configurations using gds.pipeline.list(). When a pipeline is no longer needed, you can remove it with gds.pipeline.drop().

Let's see the pipelines we currently have:

Great job! You've learned how GDS Pipelines provide a structured and reproducible way to integrate graph algorithms with machine learning workflows.

• Pipelines define a sequence of steps.
• They combine feature engineering (from existing properties and graph algorithms) with ML model training.
• They enable efficient prediction and operationalization of graph-based insights.

Keep exploring GDS to unlock more advanced graph analytics!