CODEWITHSUNDEEP
javalinear-regressionwekapmml
LexByte
LexByte

Reputation: 412

How to export Weka model made in Java to PMML format

Is there a way i can export a LinearRegression model ( build on some dataset) into a PMML format in Java?

The code so far

    DataSource source = new DataSource("house.arff");

    Instances dataset = source.getDataSet();

    Instances m_structure = new Instances(dataset, 0);

    m_structure.setClassIndex(dataset.numAttributes()-1);
    dataset.setClassIndex(dataset.numAttributes()-1);


    LinearRegression lReg = new LinearRegression();


    int m_NumClasses = dataset.numClasses();

    int class_index= dataset.classIndex();


    int nK = m_NumClasses - 1;
    int nR =  dataset.numAttributes() - 1;
    double[][] m_Par = new double[nR + 1][nK];
    String pmmlx= LogisticProducerHelper.toPMML(dataset,m_structure,m_Par,m_NumClasses);
    System.out.println(pmmlx);

This produces the following PMML file 

   <?xml version="1.0" encoding="UTF-8" standalone="yes"?>
<PMML version="4.1" xmlns="http://www.dmg.org/PMML-4_1">
    <Header copyright="WEKA">
        <Application name="WEKA" version="3.8.0"/>
    </Header>
    <DataDictionary>
        <DataField name="houseSize" optype="continuous"/>
        <DataField name="lotSize" optype="continuous"/>
        <DataField name="bedrooms" optype="continuous"/>
        <DataField name="granite" optype="continuous"/>
        <DataField name="bathroom" optype="continuous"/>
        <DataField name="sellingPrice" optype="continuous"/>
    </DataDictionary>
    <RegressionModel algorithmName="logisticRegression" functionName="classification" modelType="logisticRegression" normalizationMethod="softmax">
        <MiningSchema>
            <MiningField missingValueReplacement="3132.0" missingValueTreatment="asMean" name="houseSize" usageType="active"/>
            <MiningField missingValueReplacement="11788.142857142857" missingValueTreatment="asMean" name="lotSize" usageType="active"/>
            <MiningField missingValueReplacement="5.0" missingValueTreatment="asMean" name="bedrooms" usageType="active"/>
            <MiningField missingValueReplacement="0.42857142857142855" missingValueTreatment="asMean" name="granite" usageType="active"/>
            <MiningField missingValueReplacement="0.7142857142857143" missingValueTreatment="asMean" name="bathroom" usageType="active"/>
            <MiningField name="sellingPrice" usageType="predicted"/>
        </MiningSchema>
        <Output/>
    </RegressionModel>
</PMML>

The PMML file above cannot be used to predict an Instance because the model is not yet built.

Using the following line builds the Classifier.

lReg.buildClassifier(dataset);

So I am wondering is there a way that I can add the parameters learned by this classifier into the PMML file so it can be exported/imported easily as a already trained classifier?

Upvotes: 0

Views: 514

Answers (1)

hkn
hkn

Reputation: 1448

According to JavaDoc of LogisticProducerHelper:

Helper class for producing PMML for a Logistic classifier. Not designed to be used directly - you should call toPMML() on a trained Logistic classifier.

JavaDoc states that only Logistic classifier implements PMMLProducer.

If you use Logistic you can use logistic.toPMML(train) method.

Upvotes: 0

Related Questions