PMML 4.1 - Changes from PMML 4.0
Associations
- Clarified the meaning of the 'numberOfItems' attribute in
AssociationModel element
- Clarified that Itemset elements must appear before AssociationRule
elements
- Added 'isScorable' attribute
- Added 'leverage' and 'affinity' attributes to the AssociationRule
element
- Association Rules models can now contain the ModelVerification
element
- Added a helpful diagram
- Changed version to "4.1" from "4.0" in the example(s)
BaselineModel
- New addition to PMML 4.1 to support change detection models
BuiltinFunctions
- Defined pow(0,0) to be one
- Added built-in functions median, product, and lowercase
- Clarified the result of the if function if no else case is present
ClusteringModel
- Clarified that clusters are identified by an implicit index
- Added attribute "id" for clusters to support identifiers other than the
implicit 1-based index identifiers. The value of this attribute is used as
the output value for "predicted Value."
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
Conformance
- Added information about how deprecation is handled in PMML
- Changed all version references from "4.0" to "4.1"
DataDictionary
- The DataField element can now contain both Interval and Value elements
instead of just one
FieldScope
- New document added to collect all the rules governing field scope and
field names that were previously scattered over several documents
Functions
- Added 'mapMissingTo' and 'invalidValueTreatment' attributes to the
Apply element
GeneralRegression
- Added 'isScorable' attribute
- Added Categories element for specifying a contrast matrix
- Clarified the 'targetVariableName' attribute
- Changed version to "4.1" from "4.0" in the example(s)
GeneralStructure
- Added section explaining non-scoring models
- Added the 'mixed' mining function
- The rules about names of fields have been moved to the new field scope
document
- Updated the XML namespace to "https://www.dmg.org/PMML-4_1" from
"https://www.dmg.org/PMML-4_0"
Header
- Made 'copyright' attribute optional
- Added recommendation for the format of the 'timestamp' attribute
Interoperability
KNN
- New addition to PMML 4.1 to support k-Nearest Neighbors models
MiningSchema
- Mining fields with usage type of order can now be used for time series
models
- Mining fields are no longer required to refer to data fields, and the
rules regarding the fields that mining fields can reference have been
clarified in the FieldScope document
- The MiningSchema is the Gate Keeper
ModelExplanation
- Corrected the correlation example
- Added 'rootMeanSquaredError', 'adj-r-squared', 'sumSquaredError',
'sumSquaredRegression', 'numOfRecords', 'numOfPredictors',
'degreesOfFreedom', 'fStatistic', 'AIC', 'BIC' and 'AICc' attributes to the
PredictiveModelQuality element
ModelVerification
- Added support for zero-length strings in model verification
records
- Added an explanation of how to use model verification with Association
Rules models
MultipleModels
- Mining model segments can now contain any models allowed at the top
level of a PMML document
- Model Composition using Regression and/or DecisionTree elements has
been deprecated
- Added 'modelChain' as a possible for the 'multipleModelMethod'
attribute of the Segmentation element
- Removed LocalTransformations as a possible child element of the
Segmentation element
- Added 'isScorable' attribute
- Added a default value for the 'weight' attribute of Segment
elements
- Clarified how segments can be identified if the 'id' attribute is not
present
- Corrected the examples
- Changed version to "4.1" from "4.0" in the example(s)
NaïveBayes
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
NeuralNetwork
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
Output
- Clarified that clusters are identified by an implicit index
- Output fields can now include post-processing transformations
- Added the Decisions and Decision elements
- Leverage and affinity have been added to the possible values for
'ruleFeature' and 'rankBasis'
- Added 'segmentId' attribute to OutputField element
- Added 'transformedValue', 'decision' and 'reasonCode' to the possible
values for RESULT-FEATURE
- Deprecated feature "clusterId". Feature "entityId" should be used
instead
- Deprecated features "clusterAffinity" and "entityAffinity". The new
general feature "affinity" should be used instead
Regression
- Clarified what to do if the input to a regression table predictor is
missing
- Added 'isScorable' attribute
- Added 'name' attribute to PredictorTerm element
- Revised how to compute the probability of a target value
- Changed version to "4.1" from "4.0" in the example(s)
RuleSet
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
Scorecard
- New addition to PMML 4.1 to support Scorecard models
Sequence
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
Statistics
- Clarified what information should be in a UnivariateStats element
- Added MultivariateStats element
SupportVectorMachine
- Added 'isScorable' attribute
- Changed version to "4.1" from "4.0" in the example(s)
Targets
Taxonomy
Text
- Changed version to "4.1" from "4.0" in the example(s)
TimeSeriesModel
- Added 'isScorable' attribute
- Added support for ModelExplanation
- Removed TimeException element as a possible child of TimeSeries
element
- Corrected spelling of "seconds" in values for the TIME-ANCHOR type
- Replaced the Trend element with the Trend_ExpoSmooth element
- Added support for general polynomial exponential smoothing
- Changed version to "4.1" from "4.0" in the example(s)
Transformations
- Added 'mapMissingTo' attribute to the FieldRef element
- Most of the rules about names of DerivedField elements have been moved
to the new field scope document
TreeModel
- Added a minor clarification concerning the no true child strategy
- Added 'isScorable' attribute
- Added 'probability' attribute to the ScoreDistribution element
- Changed version to "4.1" from "4.0" in the example(s)
|