Pega^® Platform supports the import of predictive models in the Predictive Model Markup Language (PMML) format version 4.3. PMML developers can build their own models using the latest PMML format version, business users and data scientists can upload models that were built in third-party PMML tools using the latest PMML format version.

Support for PMML 4.3 on Pega Platform introduces updates to the following elements:

• General structure

  • The PROB-NUMBER type accepts scientific notation.

• Built-in functions

  The following functions were added:

  • normalCDF — Normal Cumulative Density Function.
  • normalPDF — Normal Probability Density Function.
  • stdNormalCDF — Standard Cumulative Density Function.
  • stdNormalPDF — Standards Normal Probability Density Function.
  • erf — Related Error Function.
  • normalIDF — Normal Inverse Distribution Function.
  • stdNormalIDF — Standard Normal Inverse Distribution Function.

• Output

  • The dataType attribute of the OutputField element is now required.
  • Added the isFinalResult attribute to the OutputField element.

• Model explanation

  • Multiple instances of the LiftData data element are allowed in the PredictiveModelQuality element.

• Scope of fields

  • The OutputField element with the feature="transformedValue" element can forward a reference to a derived field that is defined in the LocalTransformations element.

• Association rules

  • Added field and category attributes to the Item element.

• Neural networks

  • Added an activation function that is called rectifier.

• Support vector machines

  • Added the maxWins attribute to the SupportVectorMachineModel model.
  • Added the CategoricalPredictor element to the VectorFields element.
  • Added the CategoricalPredictor element as an alternative to the FieldRef element in the VectorFields element.

• Transformations

  • Removed the fixed method="indicator" attribute from the NormDiscrete element as it served no useful purpose. The change is only in the generated data model. The fixed method="indicator" is still supported for compatibility purposes, but triggers errors when it used in a PMML model version 4.3 or higher.

For a full list of changes in PMML 4.3, see the Data Mining Group documentation about PMML 4.3 - Changes from PMML 4.2.1.