features.md

Features

Overview

The JPMML-Evaluator library is de facto the reference implementation of the PMML specification for the Java platform.

The primary objective is to provide full compliance with 4.X versions of the PMML specification (released since 2009). The secondary objective is to provide maximum working compliance with 3.X versions of the PMML specification (released between 2004 and 2007). It means that some PMML features whose scope is limited to 3.X versions (eg. removed or deprecated in 4.X versions) may not be supported.

General structure

The JPMML-Evaluator library is hardwired to perform thorough "sanity" checking. Model evaluator classes will throw an exception when an invalid and/or unsupported PMML feature is encountered.

Data flow

• Pre-processing of active fields (aka independent variables) according to the DataDictionary and MiningSchema elements:
  • Strict data type system:
    • Except for dateDaysSince[0] and dateTimeSecondsSince[0] data types.
  • Strict operational type system.
  • Treatment of outlier, missing and/or invalid values.
• Model evaluation.
• Post-processing of target fields (aka dependent variables) according to the Targets element:
  • Rescaling and/or casting regression results.
  • Replacing a missing regression result with the default value.
  • Replacing a missing classification result with the map of prior probabilities.
• Calculation of auxiliary output fields according to the Output element:
  • Over 20 different result feature types:
    • Except for the standardError result feature.

Data manipulation

• Transformations:
  • Except for the Lag element.
• PMML built-in functions:
  • Except for erf, normalCDF, normalIDF, normalPDF, stdNormalCDF, stdNormalIDF and stdNormalPDF functions.
• PMML user-defined functions.
• Java user-defined functions.

Model types

Supported model types:

• Assocation rules (association).
• Cluster model (clustering):
  • Except for the distributionBased value of the modelClass attribute.
• General regression (regression, Cox regression, classification).
• k-Nearest neighbors (k-NN) (regression, classification, clustering).
• Naive Bayes (classification).
• Neural network (regression, classification):
  • Except for the radialBasis value of the activationFunction attribute.
• Regression (regression, classification).
• Rule set (classification).
• Scorecard (regression).
• Tree model (regression, classification):
  • Except for aggregateNodes and weightedConfidence values of the missingValueStrategy attribute.
• Support Vector Machine (SVM) (regression, classification):
  • Except for the Coefficients value of the svmRepresentation attribute.
• Ensemble model (ensembles of all of the above model types) (regression, classification, clustering).

Not yet supported model types:

• Baseline model
• Bayesian network
• Gaussian process
• Sequence rules
• Text model
• Time series model

Known limitations

• Model composition. Model composition specifies a mechanism for embedding defeatured regression and decision tree models (represented by the Regression and DecisionTree elements, respectively) into other models. Model composition was deprecated in PMML schema version 4.1.
• The ClusteringModel/CenterFields element. This element was removed in PMML schema version 3.2. PMML producers should move the list of DerivedField child elements to the ClusteringModel/LocalTransformations element, and reference them using a list of ClusteringField elements instead.
• The MiningModel/Segmentation/LocalTransformations element. This element was deprecated in PMML schema version 4.1. PMML producers should move the list of DerivedField child elements to the MiningModel/LocalTransformations element instead.
• The TableLocator element. The TableLocator element specifies a mechanism for incorporating data from external data sources (eg. CSV files, databases). The TableLocator element is simply a placeholder in PMML schema version 4.3.

Inspection API

The class model object can be inspected for unsupported PMML elements and attributes using a visitor class org.jpmml.evaluator.visitors.UnsupportedMarkupInspector (source). This visitor collects all unsupported markup as instances of org.jpmml.manager.UnsupportedMarkupException.

The class model object is safe for evaluation using the JPMML-Evaluator library if the collection of exceptions is empty:

public boolean isFullySupported(PMML pmml){
  UnsupportedMarkupInspector inspector = new UnsupportedMarkupInspector();

  // Traverse the specified class model object
  pmml.accept(inspector);

  List<UnsupportedMarkupException> exceptions = inspector.getExceptions();
  if(exceptions.isEmpty()){
    return true;
  }

  return false;
}

The visitor class traverses the class model object completely. In contrast, actual model evaluator classes traverse the class model object more or less partially, whereas every "evaluation path" is a function of the specified input data record. It follows that the collection of exceptions represents the worst case scenario. The evaluation using the JPMML-Evaluator library may succed even if this collection is not empty.