Feature engineering

Decorate a feature vector with enriched features specific to the deployed model

“Coming up with features is difficult, time-consuming, requires expert knowledge. 'Applied machine learning' is basically feature engineering.” — Prof. Andrew Ng.

Objective

Joule provides a feature engineering processor that enables users to define how features are to be created ready for predictive analytics use cases.

The processor generates for each declared feature field an engineered value. Two methods are supported:

• raw

• compute values using custom expression and plugins

On completion a feature map is generated with all the required features and placed in the StreamEvent ready for the next processor in the pipeline.

Example

feature engineering:
  name: retailProfilingFeatures
  versioned: true
  features:
    as values:
      - location_code
      - store_id

    compute:
      spend_ratio:
        scripting:
          macro:
            expression: 1 - spend/avg_spend
            language: js
            variables:
              avg_spend: 133.78
      age:
        function:
          age binning:
            source field: date_of_birth
      day:
        function:
          day-of-week transform:
            source field: date

Attributes schema

Top level attributes

Attribute	Description	Data Type	Required
name	Name feature set which is used for a predicting model	String
versioned	A boolean flag to apply a unique version identifier to the resulting feature map	Boolean Default: true
features	List of supported feature functions	List

The features attribute provide two key elements, as value and compute. Either one of the attributes must be defined.

feature engineering:
  name: retailProfilingFeatures
  versioned: true
  features:
    as values:
      - location_code
      - store_id

Attributes schema

Attribute	Description	Data Type	Required
as values	List of event fields whose value will be copied in to the feature map without any changes	List
compute	List of supported feature functions mapped to output variables to be executed using the passed event	List

feature engineering:
  ...
  features:
    as values:
      - event field1
      - event field2
      
    compute:
      output_field:
        scripting:
          ...
      other_output_field:
        function:
          plugin_name:
            ... < plugin setting > ...
            event fields:
              - f
            variables:
              varname: value

Supported feature engineering

As value

This is most basic function whereby the StreamEvent field value is copied in to the feature map.

Example

The following example will copy the location_code and store_id values directly in to the feature map.

feature engineering:
  ...
  features:
    as values:
      - location_code
      - store_id

Expression based

Joule core provides the ability to deploy declarative expressions using the custom analytics processor. This has been reused within the context of feature engineering to enable users to define custom calculations within the DSL.

Example

The following example computes per event, the spend ration based utilising a Javascript expression.

feature engineering:
  ...
  features:
    compute:
      spend_ratio:
        scripting:
          macro:
            expression: 1 - spend/avg_spend
            variables:
              avg_spend: 133.78

Custom Plugins

Developers can extend the feature engineering capabilities by extending the AbstractFeatureEngineeringFunction interface.

See CustomUserPlugin API documentation for further details.

Example

The following example computes per event, the scale price based utilising the MinMax algorithm.

This example implements the AbstractFeatureEngineeringFunction class.

feature engineering:
  ...
  features:
    compute:
      scaled_price:
        function:
          minmax scaler:
            source field: price
            variables:
              min: 10.00
              max: 12.78

Available options

Joule provides a small set of OOTB feature engineering functions.

Scripting

Define custom analytics with declarative expressions

Scaling

Normalise data with various scaling methods

Transform

Generate analytics-ready features from data

Versioning

Every feature map created is versioned using a random UUID.

The version is place directly in to the resulting map and accessed using the feature_version key.