source: branches/MetisMQI/src/main/java/weka/classifiers/meta/RealAdaBoost.java @ 37

/*
 *    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.
 */

/*
 *    RealAdaBoost.java
 *    Copyright (C) 1999, 2009 University of Waikato, Hamilton, New Zealand
 *
 */

package weka.classifiers.meta;

import weka.classifiers.Classifier;
import weka.classifiers.AbstractClassifier;
import weka.classifiers.Evaluation;
import weka.classifiers.RandomizableIteratedSingleClassifierEnhancer;
import weka.core.Capabilities;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.Option;
import weka.core.Randomizable;
import weka.core.RevisionUtils;
import weka.core.TechnicalInformation;
import weka.core.TechnicalInformationHandler;
import weka.core.Utils;
import weka.core.WeightedInstancesHandler;
import weka.core.Capabilities.Capability;
import weka.core.TechnicalInformation.Field;
import weka.core.TechnicalInformation.Type;

import java.util.Enumeration;
import java.util.Random;
import java.util.Vector;

/**
 <!-- globalinfo-start -->
 * Class for boosting a 2-class classifier using the Real Adaboost method.<br/>
 * <br/>
 * For more information, see<br/>
 * <br/>
 * J. Friedman, T. Hastie, R. Tibshirani (2000). Additive Logistic Regression: a Statistical View of Boosting. Annals of Statistics. 95(2):337-407.
 * <p/>
 <!-- globalinfo-end -->
 *
 <!-- technical-bibtex-start -->
 * BibTeX:
 * <pre>
 * &#64;article{Friedman2000,
 *    author = {J. Friedman and T. Hastie and R. Tibshirani},
 *    journal = {Annals of Statistics},
 *    number = {2},
 *    pages = {337-407},
 *    title = {Additive Logistic Regression: a Statistical View of Boosting},
 *    volume = {95},
 *    year = {2000}
 * }
 * </pre>
 * <p/>
 <!-- technical-bibtex-end -->
 *
 <!-- options-start -->
 * Valid options are: <p/>
 * 
 * <pre> -P &lt;num&gt;
 *  Percentage of weight mass to base training on.
 *  (default 100, reduce to around 90 speed up)</pre>
 * 
 * <pre> -Q
 *  Use resampling for boosting.</pre>
 * 
 * <pre> -H &lt;num&gt;
 *  Shrinkage parameter.
 *  (default 1)</pre>
 * 
 * <pre> -S &lt;num&gt;
 *  Random number seed.
 *  (default 1)</pre>
 * 
 * <pre> -I &lt;num&gt;
 *  Number of iterations.
 *  (default 10)</pre>
 * 
 * <pre> -D
 *  If set, classifier is run in debug mode and
 *  may output additional info to the console</pre>
 * 
 * <pre> -W
 *  Full name of base classifier.
 *  (default: weka.classifiers.trees.DecisionStump)</pre>
 * 
 * <pre> 
 * Options specific to classifier weka.classifiers.trees.DecisionStump:
 * </pre>
 * 
 * <pre> -D
 *  If set, classifier is run in debug mode and
 *  may output additional info to the console</pre>
 * 
 <!-- options-end -->
 *
 * Options after -- are passed to the designated classifier.<p>
 *
 * @author Eibe Frank (eibe@cs.waikato.ac.nz)
 * @author Len Trigg (trigg@cs.waikato.ac.nz)
 * @version $Revision: 6136 $ 
 */
public class RealAdaBoost 
  extends RandomizableIteratedSingleClassifierEnhancer 
  implements WeightedInstancesHandler, TechnicalInformationHandler {

  /** for serialization */
  static final long serialVersionUID = -7378109809933197974L;

  /** The number of successfully generated base classifiers. */
  protected int m_NumIterationsPerformed;

  /** Weight Threshold. The percentage of weight mass used in training */
  protected int m_WeightThreshold = 100;

  /** The value of the shrinkage parameter */
  protected double m_Shrinkage = 1;

  /** Use boosting with reweighting? */
  protected boolean m_UseResampling;
  
  /** a ZeroR model in case no model can be built from the data */
  protected Classifier m_ZeroR;

  /** Sum of weights on training data */
  protected double m_SumOfWeights;
    
  /**
   * Constructor.
   */
  public RealAdaBoost() {
    
    m_Classifier = new weka.classifiers.trees.DecisionStump();
  }
    
  /**
   * Returns a string describing classifier
   * @return a description suitable for
   * displaying in the explorer/experimenter gui
   */
  public String globalInfo() {
 
    return "Class for boosting a 2-class classifier using the Real Adaboost method.\n\n"
      + "For more information, see\n\n"
      + getTechnicalInformation().toString();
  }

  /**
   * Returns an instance of a TechnicalInformation object, containing 
   * detailed information about the technical background of this class,
   * e.g., paper reference or book this class is based on.
   * 
   * @return the technical information about this class
   */
  public TechnicalInformation getTechnicalInformation() {
    TechnicalInformation        result;

    result = new TechnicalInformation(Type.ARTICLE);
    result.setValue(Field.AUTHOR, "J. Friedman and T. Hastie and R. Tibshirani");
    result.setValue(Field.TITLE, "Additive Logistic Regression: a Statistical View of Boosting");
    result.setValue(Field.JOURNAL, "Annals of Statistics");
    result.setValue(Field.VOLUME, "95");
    result.setValue(Field.NUMBER, "2");
    result.setValue(Field.PAGES, "337-407");
    result.setValue(Field.YEAR, "2000");
    
    return result;
  }

  /**
   * String describing default classifier.
   * 
   * @return the default classifier classname
   */
  protected String defaultClassifierString() {
    
    return "weka.classifiers.trees.DecisionStump";
  }

  /**
   * Select only instances with weights that contribute to 
   * the specified quantile of the weight distribution
   *
   * @param data the input instances
   * @param quantile the specified quantile eg 0.9 to select 
   * 90% of the weight mass
   * @return the selected instances
   */
  protected Instances selectWeightQuantile(Instances data, double quantile) { 

    int numInstances = data.numInstances();
    Instances trainData = new Instances(data, numInstances);
    double [] weights = new double [numInstances];

    double sumOfWeights = 0;
    for(int i = 0; i < numInstances; i++) {
      weights[i] = data.instance(i).weight();
      sumOfWeights += weights[i];
    }
    double weightMassToSelect = sumOfWeights * quantile;
    int [] sortedIndices = Utils.sort(weights);

    // Select the instances
    sumOfWeights = 0;
    for(int i = numInstances - 1; i >= 0; i--) {
      Instance instance = (Instance)data.instance(sortedIndices[i]).copy();
      trainData.add(instance);
      sumOfWeights += weights[sortedIndices[i]];
      if ((sumOfWeights > weightMassToSelect) && 
          (i > 0) && 
          (weights[sortedIndices[i]] != weights[sortedIndices[i - 1]])) {
        break;
      }
    }
    if (m_Debug) {
      System.err.println("Selected " + trainData.numInstances()
                         + " out of " + numInstances);
    }
    return trainData;
  }

  /**
   * Returns an enumeration describing the available options.
   *
   * @return an enumeration of all the available options.
   */
  public Enumeration listOptions() {

    Vector newVector = new Vector();

    newVector.addElement(new Option(
        "\tPercentage of weight mass to base training on.\n"
        +"\t(default 100, reduce to around 90 speed up)",
        "P", 1, "-P <num>"));
    
    newVector.addElement(new Option(
        "\tUse resampling for boosting.",
        "Q", 0, "-Q"));

    newVector.addElement(new Option(
              "\tShrinkage parameter.\n"
              +"\t(default 1)",
              "H", 1, "-H <num>"));

    Enumeration enu = super.listOptions();
    while (enu.hasMoreElements()) {
      newVector.addElement(enu.nextElement());
    }
    
    return newVector.elements();
  }


  /**
   * Parses a given list of options. <p/>
   *
   <!-- options-start -->
   * Valid options are: <p/>
   * 
   * <pre> -P &lt;num&gt;
   *  Percentage of weight mass to base training on.
   *  (default 100, reduce to around 90 speed up)</pre>
   * 
   * <pre> -Q
   *  Use resampling for boosting.</pre>
   * 
   * <pre> -H &lt;num&gt;
   *  Shrinkage parameter.
   *  (default 1)</pre>
   * 
   * <pre> -S &lt;num&gt;
   *  Random number seed.
   *  (default 1)</pre>
   * 
   * <pre> -I &lt;num&gt;
   *  Number of iterations.
   *  (default 10)</pre>
   * 
   * <pre> -D
   *  If set, classifier is run in debug mode and
   *  may output additional info to the console</pre>
   * 
   * <pre> -W
   *  Full name of base classifier.
   *  (default: weka.classifiers.trees.DecisionStump)</pre>
   * 
   * <pre> 
   * Options specific to classifier weka.classifiers.trees.DecisionStump:
   * </pre>
   * 
   * <pre> -D
   *  If set, classifier is run in debug mode and
   *  may output additional info to the console</pre>
   * 
   <!-- options-end -->
   *
   * Options after -- are passed to the designated classifier.<p>
   *
   * @param options the list of options as an array of strings
   * @throws Exception if an option is not supported
   */
  public void setOptions(String[] options) throws Exception {

    String thresholdString = Utils.getOption('P', options);
    if (thresholdString.length() != 0) {
      setWeightThreshold(Integer.parseInt(thresholdString));
    } else {
      setWeightThreshold(100);
    }

    String shrinkageString = Utils.getOption('H', options);
    if (shrinkageString.length() != 0) {
      setShrinkage(new Double(shrinkageString).
        doubleValue());
    } else {
      setShrinkage(1.0);
    }
      
    setUseResampling(Utils.getFlag('Q', options));

    super.setOptions(options);
  }

  /**
   * Gets the current settings of the Classifier.
   *
   * @return an array of strings suitable for passing to setOptions
   */
  public String[] getOptions() {
    Vector        result;
    String[]      options;
    int           i;
    
    result = new Vector();

    if (getUseResampling())
      result.add("-Q");

    result.add("-P");
    result.add("" + getWeightThreshold());

    result.add("-H");
    result.add("" + getShrinkage());
    
    options = super.getOptions();
    for (i = 0; i < options.length; i++)
      result.add(options[i]);

    return (String[]) result.toArray(new String[result.size()]);
  }
  
  /**
   * Returns the tip text for this property
   * @return tip text for this property suitable for
   * displaying in the explorer/experimenter gui
   */
  public String shrinkageTipText() {
    return "Shrinkage parameter (use small value like 0.1 to reduce "
      + "overfitting).";
  }
                         
  /**
   * Get the value of Shrinkage.
   *
   * @return Value of Shrinkage.
   */
  public double getShrinkage() {
    
    return m_Shrinkage;
  }
  
  /**
   * Set the value of Shrinkage.
   *
   * @param newShrinkage Value to assign to Shrinkage.
   */
  public void setShrinkage(double newShrinkage) {
    
    m_Shrinkage = newShrinkage;
  }
  
  /**
   * Returns the tip text for this property
   * @return tip text for this property suitable for
   * displaying in the explorer/experimenter gui
   */
  public String weightThresholdTipText() {
    return "Weight threshold for weight pruning.";
  }

  /**
   * Set weight threshold
   *
   * @param threshold the percentage of weight mass used for training
   */
  public void setWeightThreshold(int threshold) {

    m_WeightThreshold = threshold;
  }

  /**
   * Get the degree of weight thresholding
   *
   * @return the percentage of weight mass used for training
   */
  public int getWeightThreshold() {

    return m_WeightThreshold;
  }
  
  /**
   * Returns the tip text for this property
   * @return tip text for this property suitable for
   * displaying in the explorer/experimenter gui
   */
  public String useResamplingTipText() {
    return "Whether resampling is used instead of reweighting.";
  }

  /**
   * Set resampling mode
   *
   * @param r true if resampling should be done
   */
  public void setUseResampling(boolean r) {

    m_UseResampling = r;
  }

  /**
   * Get whether resampling is turned on
   *
   * @return true if resampling output is on
   */
  public boolean getUseResampling() {

    return m_UseResampling;
  }

  /**
   * Returns default capabilities of the classifier.
   *
   * @return      the capabilities of this classifier
   */
  public Capabilities getCapabilities() {
    Capabilities result = super.getCapabilities();

    // class
    result.disableAllClasses();
    result.disableAllClassDependencies();
    if (super.getCapabilities().handles(Capability.BINARY_CLASS))
      result.enable(Capability.BINARY_CLASS);
    
    return result;
  }

  /**
   * Boosting method.
   *
   * @param data the training data to be used for generating the
   * boosted classifier.
   * @throws Exception if the classifier could not be built successfully
   */

  public void buildClassifier(Instances data) throws Exception {

    super.buildClassifier(data);

    // can classifier handle the data?
    getCapabilities().testWithFail(data);

    // remove instances with missing class
    data = new Instances(data);
    data.deleteWithMissingClass();
    
    m_SumOfWeights = data.sumOfWeights();

    if ((!m_UseResampling) && 
        (m_Classifier instanceof WeightedInstancesHandler)) {
      buildClassifierWithWeights(data);
    } else {
      buildClassifierUsingResampling(data);
    }
  }

  /**
   * Boosting method. Boosts using resampling
   *
   * @param data the training data to be used for generating the
   * boosted classifier.
   * @throws Exception if the classifier could not be built successfully
   */
  protected void buildClassifierUsingResampling(Instances data) 
    throws Exception {

    Instances trainData, sample, training, trainingWeightsNotNormalized;
    double sumProbs;
    int numInstances = data.numInstances();
    Random randomInstance = new Random(m_Seed);
    double minLoss = Double.MAX_VALUE;

    // Create a copy of the data so that when the weights are diddled
    // with it doesn't mess up the weights for anyone else
    trainingWeightsNotNormalized = new Instances(data, 0, numInstances);
    
    // Do boostrap iterations
    for (m_NumIterationsPerformed = -1; m_NumIterationsPerformed < m_Classifiers.length; 
         m_NumIterationsPerformed++) {
      if (m_Debug) {
        System.err.println("Training classifier " + (m_NumIterationsPerformed + 1));
      }

      training = new Instances(trainingWeightsNotNormalized);
      normalizeWeights(training, 1.0);

      // Select instances to train the classifier on
      if (m_WeightThreshold < 100) {
        trainData = selectWeightQuantile(training, 
                                         (double)m_WeightThreshold / 100);
      } else {
        trainData = new Instances(training);
      }
      
      // Resample
      double[] weights = new double[trainData.numInstances()];
      for (int i = 0; i < weights.length; i++) {
        weights[i] = trainData.instance(i).weight();
      }

      sample = trainData.resampleWithWeights(randomInstance, weights);
      
      // Build classifier
      if (m_NumIterationsPerformed == -1) {
        m_ZeroR = new weka.classifiers.rules.ZeroR();
        m_ZeroR.buildClassifier(data);
      } else {
        m_Classifiers[m_NumIterationsPerformed].buildClassifier(sample);
      }
 
      // Update instance weights
      setWeights(trainingWeightsNotNormalized, m_NumIterationsPerformed);

      // Has progress been made?
      double loss = 0;
      for (Instance inst : trainingWeightsNotNormalized) {
        loss += Math.log(inst.weight());
      }
      if (m_Debug) {
        System.err.println("Current loss on log scale: " + loss);
      }
      if ((m_NumIterationsPerformed > -1) && (loss > minLoss)) {
        if (m_Debug) {
          System.err.println("Loss has increased: bailing out.");
        }
        break;
      }
      minLoss = loss;
    }
  }

  /**
   * Sets the weights for the next iteration.
   * 
   * @param training the training instances
   * @throws Exception if something goes wrong
   */
  protected void setWeights(Instances training, int iteration) 
    throws Exception {

    for (Instance instance: training) {
      double reweight = 1;
      double prob = 1, shrinkage = m_Shrinkage;

      if (iteration == -1) {
        prob = m_ZeroR.distributionForInstance(instance)[0]; 
        shrinkage = 1.0;
      } else {
        prob = m_Classifiers[iteration].distributionForInstance(instance)[0]; 

        // Make sure that probabilities are never 0 or 1 using ad-hoc smoothing
        prob = (m_SumOfWeights * prob + 1) / (m_SumOfWeights + 2);
      }

      if (instance.classValue() == 1) {
        reweight = shrinkage * 0.5 * (Math.log(prob) - Math.log(1 - prob));
      } else {
        reweight = shrinkage * 0.5 * (Math.log(1 - prob) - Math.log(prob));
      }
      instance.setWeight(instance.weight() * Math.exp(reweight));
    }
  }

  /**
   * Normalize the weights for the next iteration.
   * 
   * @param training the training instances
   * @throws Exception if something goes wrong
   */
  protected void normalizeWeights(Instances training, double oldSumOfWeights) 
    throws Exception {

    // Renormalize weights
    double newSumOfWeights = training.sumOfWeights();
    for (Instance instance: training) {
      instance.setWeight(instance.weight() * oldSumOfWeights / newSumOfWeights);
    }
  }

  /**
   * Boosting method. Boosts any classifier that can handle weighted
   * instances.
   *
   * @param data the training data to be used for generating the
   * boosted classifier.
   * @throws Exception if the classifier could not be built successfully
   */
  protected void buildClassifierWithWeights(Instances data) 
    throws Exception {

    Instances trainData, training, trainingWeightsNotNormalized;
    int numInstances = data.numInstances();
    Random randomInstance = new Random(m_Seed);
    double minLoss = Double.MAX_VALUE;

    // Create a copy of the data so that when the weights are diddled
    // with it doesn't mess up the weights for anyone else
    trainingWeightsNotNormalized = new Instances(data, 0, numInstances);
    
    // Do boostrap iterations
    for (m_NumIterationsPerformed = -1; m_NumIterationsPerformed < m_Classifiers.length; 
         m_NumIterationsPerformed++) {
      if (m_Debug) {
        System.err.println("Training classifier " + (m_NumIterationsPerformed + 1));
      }

      training = new Instances(trainingWeightsNotNormalized);
      normalizeWeights(training, m_SumOfWeights);

      // Select instances to train the classifier on
      if (m_WeightThreshold < 100) {
        trainData = selectWeightQuantile(training, 
                                         (double)m_WeightThreshold / 100);
      } else {
        trainData = new Instances(training, 0, numInstances);
      }

      // Build classifier
      if (m_NumIterationsPerformed == -1) {
        m_ZeroR = new weka.classifiers.rules.ZeroR();
        m_ZeroR.buildClassifier(data);
      } else {
        if (m_Classifiers[m_NumIterationsPerformed] instanceof Randomizable)
          ((Randomizable) m_Classifiers[m_NumIterationsPerformed]).setSeed(randomInstance.nextInt());
        m_Classifiers[m_NumIterationsPerformed].buildClassifier(trainData);
      }

 
      // Update instance weights
      setWeights(trainingWeightsNotNormalized, m_NumIterationsPerformed);

      // Has progress been made?
      double loss = 0;
      for (Instance inst : trainingWeightsNotNormalized) {
        loss += Math.log(inst.weight());
      }
      if (m_Debug) {
        System.err.println("Current loss on log scale: " + loss);
      }
      if ((m_NumIterationsPerformed > -1) && (loss > minLoss)) {
        if (m_Debug) {
          System.err.println("Loss has increased: bailing out.");
        }
        break;
      }
      minLoss = loss;
    }
  }
  
  /**
   * Calculates the class membership probabilities for the given test instance.
   *
   * @param instance the instance to be classified
   * @return predicted class probability distribution
   * @throws Exception if instance could not be classified
   * successfully
   */
  public double [] distributionForInstance(Instance instance) 
    throws Exception {

    double [] sums = new double [instance.numClasses()]; 
    for (int i = -1; i < m_NumIterationsPerformed; i++) {
      double prob = 1, shrinkage = m_Shrinkage;
      if (i == -1) {
        prob = m_ZeroR.distributionForInstance(instance)[0]; 
        shrinkage = 1.0;
      } else {
        prob = m_Classifiers[i].distributionForInstance(instance)[0]; 
        
        // Make sure that probabilities are never 0 or 1 using ad-hoc smoothing
        prob = (m_SumOfWeights * prob + 1) / (m_SumOfWeights + 2);
      }
      sums[0] += shrinkage * 0.5 * (Math.log(prob) - Math.log(1 - prob));
    }
    sums[1] = -sums[0];
    return Utils.logs2probs(sums);
  }

  /**
   * Returns description of the boosted classifier.
   *
   * @return description of the boosted classifier as a string
   */
  public String toString() {
    
    StringBuffer text = new StringBuffer();

    if (m_ZeroR == null) {
      text.append("No model built yet.\n\n");
    } else {
      text.append("RealAdaBoost: Base classifiers: \n\n");
      text.append(m_ZeroR.toString() + "\n\n");    
      for (int i = 0; i < m_NumIterationsPerformed ; i++) {
        text.append(m_Classifiers[i].toString() + "\n\n");
      }
      text.append("Number of performed Iterations: " 
                  + m_NumIterationsPerformed + "\n");
    }

    return text.toString();
  }
  
  /**
   * Returns the revision string.
   * 
   * @return            the revision
   */
  public String getRevision() {
    return RevisionUtils.extract("$Revision: 6136 $");
  }

  /**
   * Main method for testing this class.
   *
   * @param argv the options
   */
  public static void main(String [] argv) {
    runClassifier(new RealAdaBoost(), argv);
  }
}