private SvmClassificationMethodType getClassificationMethod() {
SupportVectorMachineModel supportVectorMachineModel = getModel();
// Older versions of several popular PMML producer software are known to omit the
// classificationMethod attribute.
// The method SupportVectorMachineModel#getSvmRepresentation() replaces a missing value with the
// default value "OneAgainstAll", which may lead to incorrect behaviour.
// The workaround is to bypass this method using Java Reflection API, and infer the correct
// classification method type based on evidence.
Field field = ReflectionUtil.getField(supportVectorMachineModel, "classificationMethod");
SvmClassificationMethodType svmClassificationMethod =
ReflectionUtil.getFieldValue(field, supportVectorMachineModel);
if (svmClassificationMethod != null) {
return svmClassificationMethod;
}
List<SupportVectorMachine> supportVectorMachines =
supportVectorMachineModel.getSupportVectorMachines();
String alternateBinaryTargetCategory =
supportVectorMachineModel.getAlternateBinaryTargetCategory();
if (alternateBinaryTargetCategory != null) {
if (supportVectorMachines.size() == 1) {
SupportVectorMachine supportVectorMachine = supportVectorMachines.get(0);
String targetCategory = supportVectorMachine.getTargetCategory();
if (targetCategory != null) {
return SvmClassificationMethodType.ONE_AGAINST_ONE;
}
throw new InvalidFeatureException(supportVectorMachine);
}
throw new InvalidFeatureException(supportVectorMachineModel);
}
for (SupportVectorMachine supportVectorMachine : supportVectorMachines) {
String targetCategory = supportVectorMachine.getTargetCategory();
String alternateTargetCategory = supportVectorMachine.getAlternateTargetCategory();
if (targetCategory != null) {
if (alternateTargetCategory != null) {
return SvmClassificationMethodType.ONE_AGAINST_ONE;
}
return SvmClassificationMethodType.ONE_AGAINST_ALL;
}
throw new InvalidFeatureException(supportVectorMachine);
}
throw new InvalidFeatureException(supportVectorMachineModel);
}
private Double evaluateSupportVectorMachine(
SupportVectorMachine supportVectorMachine, double[] input) {
SupportVectorMachineModel supportVectorMachineModel = getModel();
double result = 0d;
Kernel kernel = supportVectorMachineModel.getKernel();
Coefficients coefficients = supportVectorMachine.getCoefficients();
Iterator<Coefficient> coefficientIterator = coefficients.iterator();
SupportVectors supportVectors = supportVectorMachine.getSupportVectors();
Iterator<SupportVector> supportVectorIterator = supportVectors.iterator();
Map<String, double[]> vectorMap = getVectorMap();
while (coefficientIterator.hasNext() && supportVectorIterator.hasNext()) {
Coefficient coefficient = coefficientIterator.next();
SupportVector supportVector = supportVectorIterator.next();
double[] vector = vectorMap.get(supportVector.getVectorId());
if (vector == null) {
throw new InvalidFeatureException(supportVector);
}
Double value = KernelUtil.evaluate(kernel, input, vector);
result += (coefficient.getValue() * value);
}
if (coefficientIterator.hasNext() || supportVectorIterator.hasNext()) {
throw new InvalidFeatureException(supportVectorMachine);
}
result += coefficients.getAbsoluteValue();
return result;
}
private Map<FieldName, ? extends Classification> evaluateClassification(
ModelEvaluationContext context) {
SupportVectorMachineModel supportVectorMachineModel = getModel();
List<SupportVectorMachine> supportVectorMachines =
supportVectorMachineModel.getSupportVectorMachines();
if (supportVectorMachines.size() < 1) {
throw new InvalidFeatureException(supportVectorMachineModel);
}
String alternateBinaryTargetCategory =
supportVectorMachineModel.getAlternateBinaryTargetCategory();
Classification result;
SvmClassificationMethodType svmClassificationMethod = getClassificationMethod();
switch (svmClassificationMethod) {
case ONE_AGAINST_ALL:
result = new Classification(Classification.Type.DISTANCE);
break;
case ONE_AGAINST_ONE:
result = new VoteDistribution();
break;
default:
throw new UnsupportedFeatureException(supportVectorMachineModel, svmClassificationMethod);
}
double[] input = createInput(context);
for (SupportVectorMachine supportVectorMachine : supportVectorMachines) {
String targetCategory = supportVectorMachine.getTargetCategory();
String alternateTargetCategory = supportVectorMachine.getAlternateTargetCategory();
Double value = evaluateSupportVectorMachine(supportVectorMachine, input);
switch (svmClassificationMethod) {
case ONE_AGAINST_ALL:
{
if (targetCategory == null || alternateTargetCategory != null) {
throw new InvalidFeatureException(supportVectorMachine);
}
result.put(targetCategory, value);
}
break;
case ONE_AGAINST_ONE:
if (alternateBinaryTargetCategory != null) {
if (targetCategory == null || alternateTargetCategory != null) {
throw new InvalidFeatureException(supportVectorMachine);
}
String label;
long roundedValue = Math.round(value);
// "A rounded value of 1 corresponds to the targetCategory attribute of the
// SupportVectorMachine element"
if (roundedValue == 1) {
label = targetCategory;
} else
// "A rounded value of 0 corresponds to the alternateBinaryTargetCategory attribute of
// the SupportVectorMachineModel element"
if (roundedValue == 0) {
label = alternateBinaryTargetCategory;
} else
// "The numeric prediction must be between 0 and 1"
{
throw new EvaluationException("Invalid numeric prediction " + value);
}
Double vote = result.get(label);
if (vote == null) {
vote = 0d;
}
result.put(label, (vote + 1d));
} else {
if (targetCategory == null || alternateTargetCategory == null) {
throw new InvalidFeatureException(supportVectorMachine);
}
Double threshold = supportVectorMachine.getThreshold();
if (threshold == null) {
threshold = supportVectorMachineModel.getThreshold();
}
String label;
// "If the numeric prediction is smaller than the threshold, it corresponds to the
// targetCategory attribute"
if ((value).compareTo(threshold) < 0) {
label = targetCategory;
} else {
label = alternateTargetCategory;
}
Double vote = result.get(label);
if (vote == null) {
vote = 0d;
}
result.put(label, (vote + 1d));
}
break;
default:
break;
}
}
return TargetUtil.evaluateClassification(result, context);
}
