/*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
*/
/*
* SupportVectorMachineModel.java
* Copyright (C) 2010 University of Waikato, Hamilton, New Zealand
*
*/
package weka.classifiers.pmml.consumer;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.List;
import org.w3c.dom.Element;
import org.w3c.dom.Node;
import org.w3c.dom.NodeList;
import weka.classifiers.pmml.consumer.NeuralNetwork.MiningFunction;
import weka.core.Attribute;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.RevisionUtils;
import weka.core.Utils;
import weka.core.pmml.MiningSchema;
import weka.core.pmml.TargetMetaInfo;
import weka.core.pmml.VectorDictionary;
import weka.core.pmml.VectorInstance;
import weka.gui.Logger;
/**
* Implements a PMML SupportVectorMachineModel
*
* @author Mark Hall (mhall{[at]}pentaho{[dot]}com)
* @version $Revision: 6528 $
*/
public class SupportVectorMachineModel extends PMMLClassifier
implements Serializable {
/** For serialization */
private static final long serialVersionUID = 6225095165118374296L;
/**
* Abstract base class for kernels
*/
static abstract class Kernel implements Serializable {
/** The log object to use */
protected Logger m_log = null;
protected Kernel(Logger log) {
m_log = log;
}
/** For serialization */
private static final long serialVersionUID = -6696443459968934767L;
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector instance
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public abstract double evaluate(VectorInstance x, VectorInstance y)
throws Exception;
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector (as an array of values)
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public abstract double evaluate(VectorInstance x, double[] y)
throws Exception;
/**
* Factory method returning a new Kernel constructed from
* the supplied XML element
*
* @param svmMachineModelElement the XML element containing the kernel
* @param log the logging object to use
* @return a new Kernel
* @throws Exception if something goes wrong
*/
public static Kernel getKernel(Element svmMachineModelElement,
Logger log) throws Exception {
NodeList kList = svmMachineModelElement.getElementsByTagName("LinearKernelType");
if (kList.getLength() > 0) {
return new LinearKernel(log);
}
kList = svmMachineModelElement.getElementsByTagName("PolynomialKernelType");
if (kList.getLength() > 0) {
return new PolynomialKernel((Element)kList.item(0), log);
}
kList = svmMachineModelElement.getElementsByTagName("RadialBasisKernelType");
if (kList.getLength() > 0) {
return new RadialBasisKernel((Element)kList.item(0), log);
}
kList = svmMachineModelElement.getElementsByTagName("SigmoidKernelType");
if (kList.getLength() > 0) {
return new SigmoidKernel((Element)kList.item(0), log);
}
throw new Exception("[Kernel] Can't find a kernel that I recognize!");
}
}
/**
* Subclass of Kernel implementing a simple linear (dot product) kernel
*/
static class LinearKernel extends Kernel implements Serializable {
public LinearKernel(Logger log) {
super(log);
}
public LinearKernel() {
super(null);
}
/** For serialization */
private static final long serialVersionUID = 8991716708484953837L;
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector instance
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, VectorInstance y)
throws Exception {
return x.dotProduct(y);
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector (as an array of values)
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, double[] y)
throws Exception {
return x.dotProduct(y);
}
/**
* Return a textual description of this kernel
*
* @return a string describing this kernel
*/
public String toString() {
return "Linear kernel: K(x,y) = <x,y>";
}
}
/**
* Subclass of Kernel implementing a polynomial kernel
*/
static class PolynomialKernel extends Kernel implements Serializable {
/** For serialization */
private static final long serialVersionUID = -616176630397865281L;
protected double m_gamma = 1;
protected double m_coef0 = 1;
protected double m_degree = 1;
public PolynomialKernel(Element polyNode) {
this(polyNode, null);
}
public PolynomialKernel(Element polyNode, Logger log) {
super(log);
String gammaString = polyNode.getAttribute("gamma");
if (gammaString != null && gammaString.length() > 0) {
try {
m_gamma = Double.parseDouble(gammaString);
} catch (NumberFormatException e) {
String message = "[PolynomialKernel] : WARNING, can't parse "
+ "gamma attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
String coefString = polyNode.getAttribute("coef0");
if (coefString != null && coefString.length() > 0) {
try {
m_coef0 = Double.parseDouble(coefString);
} catch (NumberFormatException e) {
String message = "[PolynomialKernel] : WARNING, can't parse "
+ "coef0 attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
String degreeString = polyNode.getAttribute("degree");
if (degreeString != null && degreeString.length() > 0) {
try {
m_degree = Double.parseDouble(degreeString);
} catch (NumberFormatException e) {
String message = "[PolynomialKernel] : WARNING, can't parse "
+ "degree attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector instance
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, VectorInstance y)
throws Exception {
double dotProd = x.dotProduct(y);
return Math.pow(m_gamma * dotProd + m_coef0, m_degree);
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector (as an array of values)
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, double[] y)
throws Exception {
double dotProd = x.dotProduct(y);
return Math.pow(m_gamma * dotProd + m_coef0, m_degree);
}
/**
* Return a textual description of this kernel
*
* @return a string describing this kernel
*/
public String toString() {
return "Polynomial kernel: K(x,y) = (" + m_gamma + " * <x,y> + "
+ m_coef0 +")^" + m_degree;
}
}
/**
* Subclass of Kernel implementing a radial basis function kernel
*/
static class RadialBasisKernel extends Kernel implements Serializable {
/** For serialization */
private static final long serialVersionUID = -3834238621822239042L;
protected double m_gamma = 1;
public RadialBasisKernel(Element radialElement) {
this(radialElement, null);
}
public RadialBasisKernel(Element radialElement, Logger log) {
super(log);
String gammaString = radialElement.getAttribute("gamma");
if (gammaString != null && gammaString.length() > 0) {
try {
m_gamma = Double.parseDouble(gammaString);
} catch (NumberFormatException e) {
String message = "[RadialBasisKernel] : WARNING, can't parse "
+ "gamma attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector instance
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, VectorInstance y)
throws Exception {
VectorInstance diff = x.subtract(y);
double result = -m_gamma * diff.dotProduct(diff);
return Math.exp(result);
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector (as an array of values)
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, double[] y)
throws Exception {
VectorInstance diff = x.subtract(y);
// System.err.println("diff: " + diff.getValues());
double result = -m_gamma * diff.dotProduct(diff);
// System.err.println("Result: " + result);
return Math.exp(result);
}
/**
* Return a textual description of this kernel
*
* @return a string describing this kernel
*/
public String toString() {
return "Radial kernel: K(x,y) = exp(-" + m_gamma + " * ||x - y||^2)";
}
}
/**
* Subclass of Kernel implementing a sigmoid function
*/
static class SigmoidKernel extends Kernel implements Serializable {
/** For serialization */
private static final long serialVersionUID = 8713475894705750117L;
protected double m_gamma = 1;
protected double m_coef0 = 1;
public SigmoidKernel(Element sigElement) {
this(sigElement, null);
}
public SigmoidKernel(Element sigElement, Logger log) {
super(log);
String gammaString = sigElement.getAttribute("gamma");
if (gammaString != null && gammaString.length() > 0) {
try {
m_gamma = Double.parseDouble(gammaString);
} catch (NumberFormatException e) {
String message = "[SigmoidKernel] : WARNING, can't parse "
+ "gamma attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
String coefString = sigElement.getAttribute("coef0");
if (coefString != null && coefString.length() > 0) {
try {
m_coef0 = Double.parseDouble(coefString);
} catch (NumberFormatException e) {
String message = "[SigmoidKernel] : WARNING, can't parse "
+ "coef0 attribute. Using default value of 1.";
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
}
}
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector instance
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, VectorInstance y)
throws Exception {
double dotProd = x.dotProduct(y);
double z = m_gamma * dotProd + m_coef0;
double a = Math.exp(z);
double b = Math.exp(-z);
return ((a - b) / (a + b));
}
/**
* Compute the result of the kernel evaluation on the supplied vectors
*
* @param x the first vector instance
* @param y the second vector (as an array of values)
* @return the result of the kernel evaluation
* @throws Exception if something goes wrong
*/
public double evaluate(VectorInstance x, double[] y)
throws Exception {
double dotProd = x.dotProduct(y);
double z = m_gamma * dotProd + m_coef0;
double a = Math.exp(z);
double b = Math.exp(-z);
return ((a - b) / (a + b));
}
/**
* Return a textual description of this kernel
*
* @return a string describing this kernel
*/
public String toString() {
return "Sigmoid kernel: K(x,y) = tanh(" + m_gamma + " * <x,y> + "
+ m_coef0 +")";
}
}
/**
* Inner class implementing a single binary (classification) SVM
*/
static class SupportVectorMachine implements Serializable {
/** For serialization */
private static final long serialVersionUID = -7650496802836815608L;
/** The target label for classification problems */
protected String m_targetCategory;
/** The index of the global alternate target label for classification */
protected int m_globalAlternateTargetCategoryIndex = -1;
/** The index of the target label for classification problems */
protected int m_targetCategoryIndex = -1;
/** PMML 4.0 - index of the alternate category for one-vs-one */
protected int m_localAlternateTargetCategoryIndex = -1;
/** PMML 4.0 - local threshold (overrides the global one, if set) */
protected double m_localThreshold = Double.MAX_VALUE; // default - not set
/** The mining schema */
protected MiningSchema m_miningSchema;
/** The log object to use (if supplied) */
protected Logger m_log;
/**
* True if this is a linear machine expressed in terms of the original
* attributes
*/
protected boolean m_coeffsOnly = false;
/** The support vectors used by this machine (if not linear with coeffs only */
protected List<VectorInstance> m_supportVectors =
new ArrayList<VectorInstance>();
/** The constant term - b */
protected double m_intercept = 0;
/** The coefficients for the vectors */
protected double[] m_coefficients;
/**
* Computes the prediction from this support vector machine. For classification,
* it fills in the appropriate element in the preds array with either a 0 or a 1.
* If the output of the machine is < 0, then a 1 is entered into the array
* element corresponding to this machine's targetCategory, otherwise a 0 is entered.
* Note that this is different to the scoring procedure in the 3.2 spec (see the comments
* in the source code of this method for more information about this).
*
* @param input the test instance to predict
* @param kernel the kernel to use
* @param vecDict the vector dictionary (null if the machine is linear and expressed
* in terms of attribute weights)
* @param preds the prediction array to fill in
* @param cMethod the classification method to use (for classification problems)
* @param globalThreshold the global threshold (used if there is no local threshold
* for this machine)
* @throws Exception if something goes wrong
*/
public void distributionForInstance(double[] input, Kernel kernel,
VectorDictionary vecDict, double[] preds, classificationMethod cMethod,
double globalThreshold)
throws Exception {
int targetIndex = 0;
if (!m_coeffsOnly) {
// get an array that only holds the values that are referenced
// by the support vectors
input = vecDict.incomingInstanceToVectorFieldVals(input);
}
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
targetIndex = m_targetCategoryIndex;
}
double result = 0;
for (int i = 0; i < m_coefficients.length; i++) {
// System.err.println("X : " + m_supportVectors.get(i).getValues());
double val = 0;
if (!m_coeffsOnly) {
val = kernel.evaluate(m_supportVectors.get(i), input);
} else {
val = input[i];
}
val *= m_coefficients[i];
// System.err.println("val " + val);
result += val;
}
result += m_intercept;
/* if (result < 0 && m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
System.err.println("[SupportVectorMachine] result ("
+ result + ") is less than zero" +
" for a nominal class value!");
result = 0;
} else if (result > 1 && m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
System.err.println("[SupportVectorMachine] result ("
+ result + ") is greater than one" +
" for a nominal class value!");
result = 1;
}
System.err.println("result " + result); */
// TODO revisit this when I actually find out what is going on with
// the Zementis model (the only easily available SVM model out there
// at this time).
if (cMethod == classificationMethod.NONE ||
m_miningSchema.getFieldsAsInstances().classAttribute().isNumeric()) {
// ----------------------------------------------------------------------
// The PMML 3.2 spec states that the output of the machine should lie
// between 0 and 1 for a binary classification case with 1 corresponding
// to the machine's targetCategory and 0 corresponding to the
// alternateBinaryTargetCategory (implying a threshold of 0.5). This seems kind of
// non standard, and indeed, in the 4.0 spec, it has changed to output < threshold
// corresponding to the machine's targetCategory and >= threshold corresponding
// to the alternateBinaryTargetCategory. What has been implemented here is
// the later with a default threshold of 0 (since the 3.2 spec doesn't have
// a way to specify a threshold). The example SVM PMML model from Zementis,
// which is PMML version 3.2, produces output between -1 and 1 (give or take).
// Implementing the 3.2 scoring mechanism as described in the spec (and truncating
// output at 0 and 1) results in all the predicted labels getting flipped on the data
// used to construct the Zementis model!
//
// April 2010 - the Zementis guys have emailed me to say that their model
// has been prepared for PMML 4.0, even though it states it is 3.2 in the
// XML.
// ----------------------------------------------------------------------
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
if (result < 0) {
preds[targetIndex] = 1;
} else {
preds[targetIndex] = 0;
}
} else {
preds[targetIndex] = result;
}
} else {
// PMML 4.0
if (cMethod == classificationMethod.ONE_AGAINST_ALL) {
// smallest value output by a machine is the predicted class
preds[targetIndex] = result;
} else {
// one-vs-one
double threshold = (m_localThreshold < Double.MAX_VALUE)
? m_localThreshold
: globalThreshold;
// vote
if (result < threshold) {
preds[targetIndex]++;
} else {
int altCat = (m_localAlternateTargetCategoryIndex != -1)
? m_localAlternateTargetCategoryIndex
: m_globalAlternateTargetCategoryIndex;
preds[altCat]++;
}
}
}
// preds[targetIndex] = result;
}
/**
* Constructs a new SupportVectorMachine from the supplied XML element
*
* @param machineElement the XML element containing the SVM
* @param miningSchema the mining schema for the PMML model
* @param dictionary the VectorDictionary from which to look up the support
* vectors used by this machine (may be null if the machine is linear and
* expressed in terms of attribute weights)
* @param svmRep the representation of this SVM (uses support vectors or is linear
* an uses attribute weights)
* @param altCategoryInd the index of the global alternateBinaryTarget (if classification)
* @param log the log object to use
* @throws Exception if something goes wrong
*/
public SupportVectorMachine(Element machineElement,
MiningSchema miningSchema, VectorDictionary dictionary, SVM_representation svmRep,
int altCategoryInd, Logger log) throws Exception {
m_miningSchema = miningSchema;
m_log = log;
String targetCat = machineElement.getAttribute("targetCategory");
if (targetCat != null && targetCat.length() > 0) {
m_targetCategory = targetCat;
Attribute classAtt = m_miningSchema.getFieldsAsInstances().classAttribute();
if (classAtt.isNominal()) {
int index = classAtt.indexOfValue(m_targetCategory);
if (index < 0) {
throw new Exception("[SupportVectorMachine] : can't find target category: "
+ m_targetCategory + " in the class attribute!");
}
m_targetCategoryIndex = index;
// now check for the PMML 4.0 alternateTargetCategory
String altTargetCat = machineElement.getAttribute("alternateTargetCategory");
if (altTargetCat != null && altTargetCat.length() > 0) {
index = classAtt.indexOfValue(altTargetCat);
if (index < 0) {
throw new Exception("[SupportVectorMachine] : can't find alternate target category: "
+ altTargetCat + " in the class attribute!");
}
m_localAlternateTargetCategoryIndex = index;
} else {
// set the global one
m_globalAlternateTargetCategoryIndex = altCategoryInd;
}
} else {
throw new Exception("[SupportVectorMachine] : target category supplied " +
"but class attribute is numeric!");
}
} else {
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
m_targetCategoryIndex = (altCategoryInd == 0)
? 1
: 0;
m_globalAlternateTargetCategoryIndex = altCategoryInd;
System.err.println("Setting target index for machine to " + m_targetCategoryIndex);
}
}
if (svmRep == SVM_representation.SUPPORT_VECTORS) {
// get the vectors
NodeList vectorsL =
machineElement.getElementsByTagName("SupportVectors");
if (vectorsL.getLength() > 0) {
Element vectors = (Element)vectorsL.item(0);
NodeList allTheVectorsL =
vectors.getElementsByTagName("SupportVector");
for (int i = 0; i < allTheVectorsL.getLength(); i++) {
Node vec = allTheVectorsL.item(i);
String vecId = ((Element)vec).getAttribute("vectorId");
VectorInstance suppV = dictionary.getVector(vecId);
if (suppV == null) {
throw new Exception("[SupportVectorMachine] : can't find " +
"vector with ID: " + vecId + " in the " +
"vector dictionary!");
}
m_supportVectors.add(suppV);
}
}
} else {
m_coeffsOnly = true;
}
// get the coefficients
NodeList coefficientsL = machineElement.getElementsByTagName("Coefficients");
// should be just one list of coefficients
if (coefficientsL.getLength() != 1) {
throw new Exception("[SupportVectorMachine] Should be just one list of " +
"coefficients per binary SVM!");
}
Element cL = (Element)coefficientsL.item(0);
String intercept = cL.getAttribute("absoluteValue");
if (intercept != null && intercept.length() > 0) {
m_intercept = Double.parseDouble(intercept);
}
// now get the individual coefficient elements
NodeList coeffL = cL.getElementsByTagName("Coefficient");
if (coeffL.getLength() == 0) {
throw new Exception("[SupportVectorMachine] No coefficients defined!");
}
m_coefficients = new double[coeffL.getLength()];
for (int i = 0; i < coeffL.getLength(); i++) {
Element coeff = ((Element) coeffL.item(i));
String val = coeff.getAttribute("value");
m_coefficients[i] = Double.parseDouble(val);
}
}
/**
* Get a textual description of this binary SVM
*
* @return a description of this SVM as a string.
*/
public String toString() {
StringBuffer temp = new StringBuffer();
temp.append("Binary SVM");
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
temp.append(" (target category = " + m_targetCategory + ")");
if (m_localAlternateTargetCategoryIndex != -1) {
temp.append("\n (alternate category = "
+ m_miningSchema.getFieldsAsInstances().classAttribute().
value(m_localAlternateTargetCategoryIndex) + ")");
}
}
temp.append("\n\n");
for (int i = 0; i < m_supportVectors.size(); i++) {
temp.append("\n" + m_coefficients[i] + " * ["
+ m_supportVectors.get(i).getValues() + " * X]");
}
if (m_intercept >= 0) {
temp.append("\n +" + m_intercept);
} else {
temp.append("\n " + m_intercept);
}
return temp.toString();
}
}
static enum SVM_representation {
SUPPORT_VECTORS,
COEFFICIENTS; // for the inputs if machine is linear and expressed in terms of the attributes
}
static enum classificationMethod {
NONE, // PMML 3.x
ONE_AGAINST_ALL, // PMML 4.0 default
ONE_AGAINST_ONE;
}
/** The mining function **/
protected MiningFunction m_functionType = MiningFunction.CLASSIFICATION;
/** The classification method (PMML 4.0) */
protected classificationMethod m_classificationMethod =
classificationMethod.NONE; // PMML 3.x (only handles binary problems)
/** The model name (if defined) */
protected String m_modelName;
/** The algorithm name (if defined) */
protected String m_algorithmName;
/** The dictionary of support vectors */
protected VectorDictionary m_vectorDictionary;
/** The kernel function to use */
protected Kernel m_kernel;
/** The individual binary SVMs */
protected List<SupportVectorMachine> m_machines =
new ArrayList<SupportVectorMachine>();
/** The other class index (in the case of a single binary SVM - PMML 3.2). */
protected int m_alternateBinaryTargetCategory = -1;
/** Do we have support vectors, or just attribute coefficients for a linear machine? */
protected SVM_representation m_svmRepresentation = SVM_representation.SUPPORT_VECTORS;
/** PMML 4.0 threshold value */
protected double m_threshold = 0;
/**
* Construct a new SupportVectorMachineModel encapsulating the information provided
* in the PMML document.
*
* @param model the SVM element from the PMML document
* @param dataDictionary the data dictionary
* @param miningSchema the mining schema
* @throws Exception if the model can't be constructed from the PMML
*/
public SupportVectorMachineModel(Element model, Instances dataDictionary,
MiningSchema miningSchema) throws Exception {
super(dataDictionary, miningSchema);
if (!getPMMLVersion().equals("3.2")) {
//TODO might have to throw an exception and only support 3.2
}
String fn = model.getAttribute("functionName");
if (fn.equals("regression")) {
m_functionType = MiningFunction.REGRESSION;
}
String modelName = model.getAttribute("modelName");
if (modelName != null && modelName.length() > 0) {
m_modelName = modelName;
}
String algoName = model.getAttribute("algorithmName");
if (algoName != null && algoName.length() > 0) {
m_algorithmName = algoName;
}
String svmRep = model.getAttribute("svmRepresentation");
if (svmRep != null && svmRep.length() > 0) {
if (svmRep.equals("Coefficients")) {
m_svmRepresentation = SVM_representation.COEFFICIENTS;
}
}
String altTargetCat = model.getAttribute("alternateBinaryTargetCategory");
if (altTargetCat != null && altTargetCat.length() > 0) {
int altTargetInd =
m_miningSchema.getFieldsAsInstances().classAttribute().indexOfValue(altTargetCat);
if (altTargetInd < 0) {
throw new Exception("[SupportVectorMachineModel] can't find alternate " +
"target value " + altTargetCat);
}
m_alternateBinaryTargetCategory = altTargetInd;
}
// PMML 4.0
String thresholdS = model.getAttribute("threshold");
if (thresholdS != null && thresholdS.length() > 0) {
m_threshold = Double.parseDouble(thresholdS);
}
// PMML 4.0
if (getPMMLVersion().startsWith("4.")) {
m_classificationMethod = classificationMethod.ONE_AGAINST_ALL; // default for PMML 4.0
}
String classificationMethodS = model.getAttribute("classificationMethod");
if (classificationMethodS != null && classificationMethodS.length()> 0) {
if (classificationMethodS.equals("OneAgainstOne")) {
m_classificationMethod = classificationMethod.ONE_AGAINST_ONE;
}
}
if (m_svmRepresentation == SVM_representation.SUPPORT_VECTORS) {
m_vectorDictionary = VectorDictionary.getVectorDictionary(model, miningSchema);
}
m_kernel = Kernel.getKernel(model, m_log);
if (m_svmRepresentation == SVM_representation.COEFFICIENTS &&
!(m_kernel instanceof LinearKernel)) {
throw new Exception("[SupportVectorMachineModel] representation is " +
"coefficients, but kernel is not linear!");
}
// Get the individual machines
NodeList machineL = model.getElementsByTagName("SupportVectorMachine");
if (machineL.getLength() == 0) {
throw new Exception("[SupportVectorMachineModel] No binary SVMs" +
" defined in model file!");
}
for (int i = 0; i < machineL.getLength(); i++) {
Node machine = machineL.item(i);
SupportVectorMachine newMach =
new SupportVectorMachine((Element)machine,
m_miningSchema, m_vectorDictionary, m_svmRepresentation,
m_alternateBinaryTargetCategory, m_log);
m_machines.add(newMach);
}
}
/**
* Classifies the given test instance. The instance has to belong to a
* dataset when it's being classified.
*
* @param inst the instance to be classified
* @return the predicted most likely class for the instance or
* Utils.missingValue() if no prediction is made
* @exception Exception if an error occurred during the prediction
*/
public double[] distributionForInstance(Instance inst) throws Exception {
if (!m_initialized) {
mapToMiningSchema(inst.dataset());
}
double[] preds = null;
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNumeric()) {
preds = new double[1];
} else {
preds = new double[m_miningSchema.getFieldsAsInstances().classAttribute().numValues()];
for (int i = 0; i < preds.length; i++) {
preds[i] = -1; // mark all entries as not calculated to begin with
}
}
double[] incoming = m_fieldsMap.instanceToSchema(inst, m_miningSchema);
boolean hasMissing = false;
for (int i = 0; i < incoming.length; i++) {
if (i != m_miningSchema.getFieldsAsInstances().classIndex() &&
Double.isNaN(incoming[i])) {
hasMissing = true;
//System.err.println("Missing value for att : " + m_miningSchema.getFieldsAsInstances().attribute(i).name());
break;
}
}
if (hasMissing) {
if (!m_miningSchema.hasTargetMetaData()) {
String message = "[SupportVectorMachineModel] WARNING: Instance to predict has missing value(s) but "
+ "there is no missing value handling meta data and no "
+ "prior probabilities/default value to fall back to. No "
+ "prediction will be made ("
+ ((m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()
|| m_miningSchema.getFieldsAsInstances().classAttribute().isString())
? "zero probabilities output)."
: "NaN output).");
if (m_log == null) {
System.err.println(message);
} else {
m_log.logMessage(message);
}
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNumeric()) {
preds[0] = Utils.missingValue();
}
return preds;
} else {
// use prior probablilities/default value
TargetMetaInfo targetData = m_miningSchema.getTargetMetaData();
if (m_miningSchema.getFieldsAsInstances().classAttribute().isNumeric()) {
preds[0] = targetData.getDefaultValue();
} else {
Instances miningSchemaI = m_miningSchema.getFieldsAsInstances();
for (int i = 0; i < miningSchemaI.classAttribute().numValues(); i++) {
preds[i] = targetData.getPriorProbability(miningSchemaI.classAttribute().value(i));
}
}
return preds;
}
} else {
for (SupportVectorMachine m : m_machines) {
m.distributionForInstance(incoming, m_kernel, m_vectorDictionary,
preds, m_classificationMethod, m_threshold);
}
}
if (m_classificationMethod != classificationMethod.NONE &&
m_miningSchema.getFieldsAsInstances().classAttribute().isNominal()) {
// PMML 4.0
if (m_classificationMethod == classificationMethod.ONE_AGAINST_ALL) {
// find the minimum value
int minI = Utils.minIndex(preds);
preds = new double[preds.length];
preds[minI] = 1.0;
} else {
// nothing to do for one-against-one - just normalize the
// votes
}
}
if (m_machines.size() == preds.length - 1) {
double total = 0;
int unset = -1;
for (int i = 0; i < preds.length; i++) {
if (preds[i] != -1) {
total += preds[i];
} else {
unset = i;
}
}
if (total > 1.0) {
throw new Exception("[SupportVectorMachineModel] total of probabilities" +
" is greater than 1!");
}
preds[unset] = 1.0 - total;
}
if (preds.length > 1) {
Utils.normalize(preds);
}
return preds;
}
public String getRevision() {
return RevisionUtils.extract("$Revision: 6528 $");
}
/**
* Get a textual description of this SupportVectorMachineModel
*
* @return a description of this SupportVectorMachineModel as a string
*/
public String toString() {
StringBuffer temp = new StringBuffer();
temp.append("PMML version " + getPMMLVersion());
if (!getCreatorApplication().equals("?")) {
temp.append("\nApplication: " + getCreatorApplication());
}
temp.append("\nPMML Model: Support Vector Machine Model");
temp.append("\n\n");
temp.append(m_miningSchema);
temp.append("Kernel: \n\t");
temp.append(m_kernel);
temp.append("\n");
if (m_classificationMethod != classificationMethod.NONE) {
temp.append("Multi-class classifcation using ");
if (m_classificationMethod == classificationMethod.ONE_AGAINST_ALL) {
temp.append("one-against-all");
} else {
temp.append("one-against-one");
}
temp.append("\n\n");
}
for (SupportVectorMachine v : m_machines) {
temp.append("\n" + v);
}
return temp.toString();
}
}