Add support for post-processing of model output

[latenighthacks]

Hi Villu, thank you for your excellent work on JPMML.

I have a multiple linear regression model in R that uses two features and predicts a response variable. Due to a project constraint, I must actually train the model to predict log10(response), so that the regression minimizes the RMSLE. However, the model should still output a prediction for the untransformed response.

My understanding is that JPMML-Evaluator can handle such post-processing of model outputs, but I have not been able to find any way to include post-processing instructions when I use r2pmml.

Would it be possible to include some way of post-processing outputs to apply a simple transformation? Or, if that is not feasible right now, would you recommend simply editing the XML by hand to add the transformation?

Thanks in advance for your time!

[vruusmann]

Related to: jpmml/jpmml-r#7

I must train the model to predict log10(response). However, the model should still output a prediction for the untransformed response.

The (multiple-) linear regression model is encoded using the RegressionModel element. The simplest way to "undo" the log transform on the target variable would be to specify RegressionModel@normalizationMethod="exp" (by default, it should be none). See this page (and scroll down to the "Valid combinations" section): http://dmg.org/pmml/v4-3/Regression.html#xsdType_REGRESSIONNORMALIZATIONMETHOD

[latenighthacks]

Hi Villu,

Thanks for your quick reply! I edited the XML for the model in R using

r2pmml(m, "model.pmml")
modelXML <- xmlTreeParse("model.pmml", useInternalNodes = TRUE)
add_normalization <- function(xmlNode) { xmlAttrs(xmlNode) <- c(normalizationMethod = "exp") }
xpathApply(modelXML, 
                 "/jpmml:PMML/jpmml:RegressionModel", 
                 add_normalization, 
                 namespaces = c(jpmml = xmlNamespaceDefinitions(modelXML, simplify = T)))
saveXML(modelXML, "model.pmml" indent = TRUE)

and when we run the model in JPMML-Evaluator, the model output is transformed (as expected).

Thanks very much for your help. I will close the issue.