PMML 3.2 - Changes from PMML 3.1
Associations
- no changes
BuiltinFunctions
- added computational examples for dataDaysSinceYear, dateSecondsSinceYear and dateSecondsSinceMidnight
- removed broken link to POSIX standard excerpt
- clarified results of builtins in case of missing input values
ClusteringModel
- Removed CenterFields and integrated functionality into CluseringFields
Conformance
- Improved the examples of aspects of the standard that cannot be enforced by the XSD [Rick Pechter]
DataDictionary
- clarified that aYear in dateDaysSince[aYear] and dateTimeSecondsSince[aYear] is not a variable
- ordinal fields with string values do not have a default ordering anymore
- Added example of a cyclic field
- Clarified order definition for ordinal fields
Functions
- Clarified that user-defined functions must be unique
GeneralRegression
- Added model type "generalizedLinear" and an example with a description of the scoring procedure.
- Renamed the enum LINK-FUNCTION to CUMULATIVE-LINK-FUNCTION.
- Added new enum LINK-FUNCTION.
- Added new attributes to the model element:
- distribution
- distParameter
- linkFunction
- linkParameter
- offsetVariable
- offsetValue
- trialsVariable
- trialsValue
- Added optional attribute "type" to PCovMatrix with possible values "model" and "robust".
- Corrected omissions.
- Added links to element descriptions inside the document.
GeneralStructure
- SparseArrays: Have separate vector elements INT-Elements and REAL-Elements to avoid XSD validation problems
- Added section regarding ties
Header
- no changes
MiningSchema
- clarified usage of displayName in DataDictionary
- clarified applicability of outliers
- declared predicted fields to be not mandatory
ModelComposition
- no changes
ModelVerification
- clarified relation to output and predicted fields
NaïveBayes
- no changes
NeuralNetwork
- added back altitude to Neuron
Output
- clarified requirements in case of multiple predicted fields.
- clarified that value refers to the displayValue if applicable.
- add table that specifies which outputs are allowed for each modeltype
Regression
- no changes
RuleSet
- no changes
Sequence
- no changes
Statistics
- added examples for UnivariateStats and Partition
SupportVectorMachine
- no changes
Targets
- Added examples for negative values for floor and ceiling
- Added Partition to TargetValue for classlabel-specific statistics
- Renamed rawDataValue to displayValue and redefined scope to be consistent with other places like DataDictionary that also have a value and displayValue.
- MiningFields of type predicted are no longer declared as optional.
Taxonomy
- no changes
Text
- no changes
Transformations
- LinearNorm is always required - removed contradicting sentence that implied otherwise
- clarified outlier treatment for NormContinuous
- Added mapMissingTo to NormContinuous
- Clarify behaviour for missing values in groupField in Aggregate
- Remove thermometer encoding in NormDiscrete
- Added new attribute outliers
TreeModel
- Clarify that only imediate children can be referenced in defaultChild
- Attribute score in Node is not required in all cases anymore
- Added Partition to Node for node-specific statistics
- Added new missingValueStrategy aggregateNodes