pmml (version 1.3)

addLT: Combines a given LocalTransformations Node with a preexisting PMML model

Description

This function is called by the 'pmml' function when adding a given new LocalTransformations node to a pre-existing PMML model.

Usage

addLT(xmlmodel=NULL, transforms=NULL)

Arguments

xmlmodel

the PMML model in as a XMLNode object.

transforms

The wrapper object obtained by using the WrapData function from the 'pmmlTransformations' package. This object contains the information on all the transformations to be performed.

Value

• An object of class XMLNode as that defined by the XML package. This represents the top level, or root node, of the XML document and is of type PMML. It can be written to file with saveXML.

Details

This is an internal function that takes as an input a PMML model as a XMLNode object and a 'LocalTransformations' element as a XMLNode object. These are the formats in which models are output when using the 'pmml' function on the R models supported by the pmml package. This function checks to see if any of the variables used in the model are actually derived from another field described in the 'LocalTransformations' element; this is the PMML element that is included in the wrapper object passed as the second argument. It then replaces the fields in the 'DataDictionary' and the 'MiningSchema' elements with the original fields from which they were derived from and adds the 'LocalTransformations' element to the existing PMML file. If a 'LocalTransformations' element already exists, the new transformations are appended to it.

It further checks the possible categories of the fields already derived and adds any new possible values. It also checks for the correct data types of the new fields. The result is a new consistent model file in PMML format which would appear as if it was made from the original variables directly. The inputs are the original variables of the right data types and the right categories, the right number of fields with no repetition and the appropriate transformations. The functionality achieved by 'addLT' is described in the book 'PMML in Action (Second Edition)' referenced to in this document (refer to Chapter 5 entitled 'The Making of a Predictive Solution').

The scenario implemented here is that the model uses fields which are later realized to be derivable from other more basic fields. The only condition is that the transformations may not contain an arbitrary transform; all transforms defined must either directly or eventually via a transformation chain end up as a field presently used by the model. Without remaking the entire model, this function is used to update the model. Consider the example of the iris data set. Some possible use cases are: 1) many different variables are actually derived from the same original field. The variable 'Species' may be mapped to various other values of a derived field 'd_Species'. Further, the same variable 'Species' may be discretized to 3 numerical variables: 'Species_setosa', 'Species_versicolor' and 'Species_virginica'. Hence multiple fields are derived from the same original field. 2) The field used in the model is derived from a chain of transformations. The example shows that 'Sepal.Length' is first derived into 'd1' and then 'd1' is derived to 'dd1'. The final field 'dd1' is hence derived from a chain of transformations starting from 'Sepal.Length'.

It is a very lengthy and complicated process to combine two models; besides just replacing the variables, change in variable data types, change in variable statistics and change in variable categories must be correctly implemented. This is not easily done by hand and this function does all this automatically for the user.

References

PMML home page: http://www.dmg.org

A. Guazzelli, W. Lin, T. Jena (2012), PMML in Action: Unleashing the Power of Open Standards for Data Mining and Predictive Analytics. CreativeSpace (Second Edition - Available on Amazon.com - http://www.amazon.com/dp/1470003244.

A. Guazzelli, M. Zeller, W. Lin, G. Williams (2009), PMML: An Open Standard for Sharing Models. The R journal, Volume 1/1, 60-65

R pmmlTransformations package: http://cran.r-project.org/web/packages/pmmlTransformations/index.html

See Also

pmml.

Examples

## Not run:
# The following commands can be run from R given a text file "class_table.csv" which, 
# for example, may contain the 2 lines:
# setosa,1
# versicolor,2
#
# Naturally, the file does not contain the hash character. This file contains information 
# which maps the input field so that its value 'setosa' is mapped to the value 1 and its 
# value 'versicolor' is mapped to 2. See the manual of the 'pmmlTransformations' package for 
# details

# Create a pmmlTransformations object
library(pmmlTransformations)
irisBox <- WrapData(iris)

# Transform the 'Sepal.Length' variable to 'dd1' via a chain
irisBox <- ZScoreXform(irisBox,xformInfo = "Sepal.Length->d1")
irisBox <- MinMaxXform(irisBox,xformInfo = "d1->dd1")

#transform the same variable 'class' to many different derived variables
irisBox <- MapXform(irisBox,xformInfo = "[Species->d_Species][string->double]",
table="class_table3.csv",defaultValue="-1",mapMissingTo="1")

irisBox <- NormDiscreteXform(irisBox,inputVar="Species")

# make a clustering model calculating 'Species' using the derived fields, then convert 
# to PMML format
fit <- kmeans(irisBox$data[,c(6:11)],5)
pmml_fit <- pmml(fit, predictedField="Species")

# now we imagine that we realize the fields we used were actually derived from the basic iris
# data set. First we put the transformation information in PMML format.
pmml_trans <- pmml(NULL, transforms=irisBox)

# Finally, we can see the combined PMML file
pmml_combined <- pmml(pmml_fit, transforms=pmml_trans)

## End(Not run)