Pega^® Platform uses a specific implementation of the PMML format, which means that some of the PMML features and models are not supported in the Predictive Model rule. PMML developers must know the supported PMML versions and models, as well as the unsupported models and features, when building PMML models for Pega Platform. Knowing these limitations prevents issues that might occur when you upload the PMML file.

Supported PMML versions

• 3.0
• 3.1
• 3.2
• 4.0
• 4.1
• 4.2
• 4.2.1
• 4.3

Supported models

• Cluster model
• General regression
• Mining (multiple/ensemble) model
• Neural Network
• k-Nearest Neighbors
• Naive Bayes
• Ruleset
• Regression
• Support Vector Machine
• Scorecard
• Tree

Unsupported models

• Association rules
• Baseline
• Bayesian Network
• Ensemble (Mining/Many-in-one) models that contain composite embedded models
• Gaussian Process
• Sequence
• Text
• Time series

Limitations

• Cluster models
  • The kind attribute of the ComparisonMeasureelement can be set only to distance or similarity.
• General regression
  • If the functionName attribute of the GeneralRegressionModel element is regression, the model must have exactly one PPMatrix. Running a model with no or multiple instances of PMMatrix results in errors.
  • The multinomialLogistic, ordinalMultinomial, and CoxRegression algorithms are not supported for the regression mining function.
  • The regression, general_linear, and CoxRegression algorithms are not supported for the classification mining function.
• Ensemble - In PMML, XGBoost and Random forest models are defined as Ensemble models.
  • Model composition and aggregation methods that you can use are:
    • Mining function regression - median.
    • Mining function classification - average, weightedAverage, median, max.
• k-Nearest Neighbors
  • The opType attribute of the input field (DataField) can be set only to continuous or categorical.
  • The kind attribute of the ComparisonMeasure element can be set only to distance or similarity.
• Naive Bayes - Supports only classificationmining function.
• Neural Network - You can only use regression or classification mining function.
• Regression
  • If the functionName attribute of the RegressionModel element has the value regression, the normalizationMethod attribute can have only one of the following values: none, softmax, logit, exp.
  • If the functionName attribute of the RegressionModel element has the value classification, the normalizationMethod attribute can have only one of the following values: none, softmax, logit, loglog, cloglog.
• Scorecard
  • In Pega Platform implementation of the PMML standard, the functionName attribute is mandatory for the ScorecardModel element.
• Support Vector Machine
  • The svmRepresentation attribute is mandatory for the SupportVectorMachineModel element.
  • The functionName attribute for the SupportVectorMachineModel element cannot be empty and has to be set to regression or classification.
  • The probability attribute value is not supported for the resultFeature attribute in the Output element.
• Tree
  • The functionName attribute for the TreeModel element cannot be empty, and has to be set to regression or classification.