vruusmann commented on a change in pull request #23868: [SPARK-26966][ML]
Update to JPMML 1.4.8
URL: https://github.com/apache/spark/pull/23868#discussion_r259231996
##########
File path:
mllib/src/main/scala/org/apache/spark/mllib/pmml/export/PMMLModelExportFactory.scala
##########
@@ -44,12 +44,12 @@ private[mllib] object PMMLModelExportFactory {
new GeneralizedLinearPMMLModelExport(lasso, "lasso regression")
case svm: SVMModel =>
new BinaryClassificationPMMLModelExport(
- svm, "linear SVM", RegressionNormalizationMethodType.NONE,
+ svm, "linear SVM", RegressionModel.NormalizationMethod.NONE,
Review comment:
`RegressionModel` is a top-level PMML model element for representing all
"dot product"-type models; the real mining function type (regression vs
classification) is specified using the `functionName` attribute. So,
`RegressionModel@functionName="classification"` is how communicates that this
model element is encoding a classifier-type function.
Based on my experience with implementing Apache Spark ML-to-PMML converters
(https://github.com/jpmml/jpmml-sparkml), then many non-decision tree based
Apache Spark ML model classes (eg. `NaiveBayesClassifier`, `LinearSVC`) are
based on the "dot product" business logic, and therefore reducible to the
`RegressionModel` element. In other words, there is no point in using more
complex PMML model elements such as `NaiveBayesModel` or
`SupportVectorMachineModel`, when the simplest `RegressionModel` element will
be able to capture everything.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
