Customizing the "missing value handling"-mode of models about jpmml-sparkml HOT 2 OPEN

Apache Spark ML decision tree models do not support such "execution flow" customizations. Therefore, it is impossible to generate the requested PMML markup automatically.

Possible workaround:

1. Generate a "raw" PMML class model object by invoking ConverterUtil#toPMML(StructType, PipelineModel).
2. Apply a (list of-) post-processing Visitor(s) to it.
3. Save the "post-processed" PMML class model object to a file.

In your case, the Visitor needs to be modifying TreeModel elements (of the top-level MiningModel element). Here's a starting point:

PMML pmml = ConverterUtil.toPMML(...);

Visitor treeModelUpdater = new new AbstractVisitor(){

	@Override
	public VisitorAction visit(TreeModel treeModel){
		treeModel.setMissingValueStrategy(MissingValueStrategy.NULL_PREDICTION);
		treeModel.setMissingValuePenalty(0d);
		treeModel.setNoTrueChildStrategy(NoTrueChildStrategy.RETURN_NULL_PREDICTION);
	
		return VisitorAction.CONTINUE;
	}
};
treeModelUpdater.applyTo(pmml);

from jpmml-sparkml.

Related to #14

It should be possible to toggle PMML converters between two modes:

1. "Missing value"-friendly (a missing input leads to a missing prediction).
2. "Missing value"-hostile (a missing input raises an error, or leads to a non-missing default prediction).

from jpmml-sparkml.