Use Azure Machine Learning-based models

The unified data in Dynamics 365 Customer Insights is a source for building machine learning models that can generate additional business insights. Customer Insights integrates with Azure Machine Learning to use your own custom models.

Prerequisites

• Access to Customer Insights
• Active Azure Enterprise subscription
• Unified customer profiles
• Entity export to Azure Blob storage configured

Set up Azure Machine Learning workspace

1. See create a Azure Machine Learning workspace for different options to create the workspace. For best performance, create the workspace in an Azure region that is geographically closest to your Customer Insights environment.

2. Access your workspace through the Azure Machine Learning Studio. There are several ways to interact with your workspace.

Work with Azure Machine Learning designer

Azure Machine Learning designer provides a visual canvas where you can drag and drop datasets and modules. A batch pipeline created from the designer can be integrated into Customer Insights if they are configured accordingly.

Working with Azure Machine Learning SDK

Data scientists and AI developers use the Azure Machine Learning SDK to build machine learning workflows. Currently, models trained using the SDK can't be integrated directly with Customer Insights. A batch inference pipeline that consumes that model is required for integration with Customer Insights.

Batch pipeline requirements to integrate with Customer Insights

Dataset Configuration

You need to create datasets to use entity data from Customer Insights to your batch inference pipeline. These datasets need to be registered in the workspace. Currently, we only support tabular datasets in .csv format. The datasets that correspond to entity data need to be parameterized as a pipeline parameter.

• Dataset parameters in Designer

  In the designer, open Select Columns in Dataset and select Set as pipeline parameter where you provide a name for the parameter.

  

• Dataset parameter in SDK (Python)

  HotelStayActivity_dataset = Dataset.get_by_name(ws, name='Hotel Stay Activity Data')
  HotelStayActivity_pipeline_param = PipelineParameter(name="HotelStayActivity_pipeline_param", default_value=HotelStayActivity_dataset)
  HotelStayActivity_ds_consumption = DatasetConsumptionConfig("HotelStayActivity_dataset", HotelStayActivity_pipeline_param)
  

Batch inference pipeline

• In the designer, a training pipeline can be used to create or update an inference pipeline. Currently, only batch inference pipelines are supported.

• Using the SDK, you can publish the pipeline to an endpoint. Currently, Customer Insights integrates with the default pipeline in a batch pipeline endpoint in the Machine Learning workspace.

  published_pipeline = pipeline.publish(name="ChurnInferencePipeline", description="Published Churn Inference pipeline")
  pipeline_endpoint = PipelineEndpoint.get(workspace=ws, name="ChurnPipelineEndpoint") 
  pipeline_endpoint.add_default(pipeline=published_pipeline)
  

Import pipeline data into Customer Insights

• The designer provides the Export Data module that allows the output of a pipeline to be exported to Azure storage. Currently, the module must use the datastore type Azure Blob Storage and parameterize the Datastore and relative Path. Customer Insights overrides both these parameters during pipeline execution with a datastore and path that is accessible to the product.

  

• When writing the inference output using code, you can upload the output to the a path within a registered datastore in the workspace. If the path and datastore are parameterized in the pipeline, Customer insights will be able to read and import the inference output. Currently, a single tabular output in csv format is supported. The path must include the directory and filename.

  # In Pipeline setup script
      OutputPathParameter = PipelineParameter(name="output_path", default_value="HotelChurnOutput/HotelChurnOutput.csv")
      OutputDatastoreParameter = PipelineParameter(name="output_datastore", default_value="workspaceblobstore")
  ...
  # In pipeline execution script
      run = Run.get_context()
      ws = run.experiment.workspace
      datastore = Datastore.get(ws, output_datastore) # output_datastore is parameterized
      directory_name =  os.path.dirname(output_path)  # output_path is parameterized.
  
      # Datastore.upload() or Dataset.File.upload_directory() are supported methods to uplaod the data
      # datastore.upload(src_dir=<<working directory>>, target_path=directory_name, overwrite=False, show_progress=True)
      output_dataset = Dataset.File.upload_directory(src_dir=<<working directory>>, target = (datastore, directory_name)) # Remove trailing "/" from directory_name