bayesianlogisticregression.java

bayes算法在java环境下的开发和实现过程

/*
 *    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.
 */

/*
 *    BayesianLogisticRegression.java
 *    Copyright (C) 2008 Illinois Institute of Technology
 *
 */

package weka.classifiers.bayes;

import weka.classifiers.Classifier;
import weka.classifiers.bayes.blr.GaussianPriorImpl;
import weka.classifiers.bayes.blr.LaplacePriorImpl;
import weka.classifiers.bayes.blr.Prior;
import weka.core.Attribute;
import weka.core.Capabilities;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.Option;
import weka.core.OptionHandler;
import weka.core.RevisionUtils;
import weka.core.SelectedTag;
import weka.core.SerializedObject;
import weka.core.Tag;
import weka.core.TechnicalInformation;
import weka.core.TechnicalInformationHandler;
import weka.core.Utils;
import weka.core.Capabilities.Capability;
import weka.core.TechnicalInformation.Field;
import weka.core.TechnicalInformation.Type;
import weka.filters.Filter;
import weka.filters.unsupervised.attribute.Normalize;

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

/**
 <!-- globalinfo-start -->
 * Implements Bayesian Logistic Regression for both Gaussian and Laplace Priors.<br/>
 * <br/>
 * For more information, see<br/>
 * <br/>
 * Alexander Genkin, David D. Lewis, David Madigan (2004). Large-scale bayesian logistic regression for text categorization. URL http://www.stat.rutgers.edu/~madigan/PAPERS/shortFat-v3a.pdf.
 * <p/>
 <!-- globalinfo-end -->
 * 
 <!-- technical-bibtex-start -->
 * BibTeX:
 * <pre>
 * &#64;techreport{Genkin2004,
 *    author = {Alexander Genkin and David D. Lewis and David Madigan},
 *    institution = {DIMACS},
 *    title = {Large-scale bayesian logistic regression for text categorization},
 *    year = {2004},
 *    URL = {http://www.stat.rutgers.edu/\~madigan/PAPERS/shortFat-v3a.pdf}
 * }
 * </pre>
 * <p/>
 <!-- technical-bibtex-end -->
 *
 *
 *  @author Navendu Garg (gargnav at iit dot edu)
 *  @version $Revision: 1.3 $
 */
public class BayesianLogisticRegression extends Classifier
  implements OptionHandler, TechnicalInformationHandler {
  
  static final long serialVersionUID = -8013478897911757631L;

  /** Log-likelihood values to be used to choose the best hyperparameter. */
  public static double[] LogLikelihood;

  /** Set of values to be used as hyperparameter values during Cross-Validation. */
  public static double[] InputHyperparameterValues;

  /** DEBUG Mode*/
  boolean debug = false;

  /** Choose whether to normalize data or not */
  public boolean NormalizeData = false;

  /** Tolerance criteria for the stopping criterion. */
  public double Tolerance = 0.0005;

  /** Threshold for binary classification of probabilisitic estimate*/
  public double Threshold = 0.5;

  /** Distributions available */
  public static final int GAUSSIAN = 1;
  public static final int LAPLACIAN = 2;
  
  public static final Tag[] TAGS_PRIOR = {
    new Tag(GAUSSIAN, "Gaussian"),
    new Tag(LAPLACIAN, "Laplacian")
  };

  /** Distribution Prior class */
  public int PriorClass = GAUSSIAN;

  /** NumFolds for CV based Hyperparameters selection*/
  public int NumFolds = 2;

  /** Methods for selecting the hyperparameter value */
  public static final int NORM_BASED = 1;
  public static final int CV_BASED = 2;
  public static final int SPECIFIC_VALUE = 3;

  public static final Tag[] TAGS_HYPER_METHOD = {
    new Tag(NORM_BASED, "Norm-based"),
    new Tag(CV_BASED, "CV-based"),
    new Tag(SPECIFIC_VALUE, "Specific value")
  };

  /** Hyperparameter selection method */
  public int HyperparameterSelection = NORM_BASED;

  /** The class index from the training data */
  public int ClassIndex = -1;

  /** Best hyperparameter for test phase */
  public double HyperparameterValue = 0.27;

  /** CV Hyperparameter Range */
  public String HyperparameterRange = "R:0.01-316,3.16";

  /** Maximum number of iterations */
  public int maxIterations = 100;

  /**Iteration counter */
  public int iterationCounter = 0;

  /** Array for storing coefficients of Bayesian regression model. */
  public double[] BetaVector;

  /** Array to store Regression Coefficient updates. */
  public double[] DeltaBeta;

  /**        Trust Region Radius Update*/
  public double[] DeltaUpdate;

  /** Trust Region Radius*/
  public double[] Delta;

  /**  Array to store Hyperparameter values for each feature. */
  public double[] Hyperparameters;

  /** R(i)= BetaVector X x(i) X y(i).
   * This an intermediate value with respect to vector BETA, input values and corresponding class labels*/
  public double[] R;

  /** This vector is used to store the increments on the R(i). It is also used to determining the stopping criterion.*/
  public double[] DeltaR;

  /**
   * This variable is used to keep track of change in
   * the value of delta summation of r(i).
   */
  public double Change;

  /**
   * Bayesian Logistic Regression returns the probability of a given instance will belong to a certain
   * class (p(y=+1|Beta,X). To obtain a binary value the Threshold value is used.
   * <pre>
   * p(y=+1|Beta,X)>Threshold ? 1 : -1
   * </pre>
   */

  /** Filter interface used to point to weka.filters.unsupervised.attribute.Normalize object
   *
   */
  public Filter m_Filter;

  /** Dataset provided to do Training/Test set.*/
  protected Instances m_Instances;

  /**        Prior class object interface*/
  protected Prior m_PriorUpdate;

  public String globalInfo() {
    return "Implements Bayesian Logistic Regression "
      + "for both Gaussian and Laplace Priors.\n\n"
      + "For more information, see\n\n"
      + getTechnicalInformation();
  }

  /**
   * <pre>
   * (1)Initialize m_Beta[j] to 0.
   * (2)Initialize m_DeltaUpdate[j].
   * </pre>
   *
   * */
  public void initialize() throws Exception {
    int numOfAttributes;
    int numOfInstances;
    int i;
    int j;

    Change = 0.0;

    //Manipulate Data
    if (NormalizeData) {
      m_Filter = new Normalize();
      m_Filter.setInputFormat(m_Instances);
      m_Instances = Filter.useFilter(m_Instances, m_Filter);
    }

    //Set the intecept coefficient.
    Attribute att = new Attribute("(intercept)");
    Instance instance;

    m_Instances.insertAttributeAt(att, 0);

    for (i = 0; i < m_Instances.numInstances(); i++) {
      instance = m_Instances.instance(i);
      instance.setValue(0, 1.0);
    }

    //Get the number of attributes
    numOfAttributes = m_Instances.numAttributes();
    numOfInstances = m_Instances.numInstances();
    ClassIndex = m_Instances.classIndex();
    iterationCounter = 0;

    //Initialize Arrays.
    switch (HyperparameterSelection) {
    case 1:
      HyperparameterValue = normBasedHyperParameter();

      if (debug) {
        System.out.println("Norm-based Hyperparameter: " + HyperparameterValue);
      }

      break;

    case 2:
      HyperparameterValue = CVBasedHyperparameter();

      if (debug) {
        System.out.println("CV-based Hyperparameter: " + HyperparameterValue);
      }

      break;
    }

    BetaVector = new double[numOfAttributes];
    Delta = new double[numOfAttributes];
    DeltaBeta = new double[numOfAttributes];
    Hyperparameters = new double[numOfAttributes];
    DeltaUpdate = new double[numOfAttributes];

    for (j = 0; j < numOfAttributes; j++) {
      BetaVector[j] = 0.0;
      Delta[j] = 1.0;
      DeltaBeta[j] = 0.0;
      DeltaUpdate[j] = 0.0;

      //TODO: Change the way it takes values.
      Hyperparameters[j] = HyperparameterValue;
    }

    DeltaR = new double[numOfInstances];
    R = new double[numOfInstances];

    for (i = 0; i < numOfInstances; i++) {
      DeltaR[i] = 0.0;
      R[i] = 0.0;
    }

    //Set the Prior interface to the appropriate prior implementation.
    if (PriorClass == GAUSSIAN) {
      m_PriorUpdate = new GaussianPriorImpl();
    } else {
      m_PriorUpdate = new LaplacePriorImpl();
    }
  }

  /**
   * This method tests what kind of data this classifier can handle.
   * return Capabilities
   */
  public Capabilities getCapabilities() {
    Capabilities result = super.getCapabilities();

    // attributes
    result.enable(Capability.NUMERIC_ATTRIBUTES);

    result.enable(Capability.BINARY_ATTRIBUTES);

    // class
    result.enable(Capability.BINARY_CLASS);

    // instances
    result.setMinimumNumberInstances(0);

    return result;
  }

  /**
   * <ul>
   *         <li>(1) Set the data to the class attribute m_Instances.</li>
   *  <li>(2)Call the method initialize() to initialize the values.</li>
   * </ul>
   *        @param data training data
   *        @exception Exception if classifier can't be built successfully.
   */
  public void buildClassifier(Instances data) throws Exception {
    Instance instance;
    int i;
    int j;

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

    //(1) Set the data to the class attribute m_Instances.
    m_Instances = new Instances(data);

    //(2)Call the method initialize() to initialize the values.
    initialize();

    do {
      //Compute the prior Trust Region Radius Update;
      for (j = 0; j < m_Instances.numAttributes(); j++) {
        if (j != ClassIndex) {
          DeltaUpdate[j] = m_PriorUpdate.update(j, m_Instances, BetaVector[j],
              Hyperparameters[j], R, Delta[j]);
          //limit step to trust region.
          DeltaBeta[j] = Math.min(Math.max(DeltaUpdate[j], 0 - Delta[j]),
              Delta[j]);

          //Update the 
          for (i = 0; i < m_Instances.numInstances(); i++) {
            instance = m_Instances.instance(i);

            if (instance.value(j) != 0) {
              DeltaR[i] = DeltaBeta[j] * instance.value(j) * classSgn(instance.classValue());
              R[i] += DeltaR[i];
            }
          }

          //Updated Beta values.
          BetaVector[j] += DeltaBeta[j];

          //Update size of trust region.
          Delta[j] = Math.max(2 * Math.abs(DeltaBeta[j]), Delta[j] / 2.0);
        }
      }
    } while (!stoppingCriterion());

    m_PriorUpdate.computelogLikelihood(BetaVector, m_Instances);
    m_PriorUpdate.computePenalty(BetaVector, Hyperparameters);
  }

  /**
   * This class is used to mask the internal class labels.
   *
   * @param value internal class label
   * @return
   * <pre>
   * <ul><li>
   *  -1 for internal class label 0
   *  </li>
   *  <li>
   *  +1 for internal class label 1
   *  </li>
   *  </ul>
   *  </pre>
   */
  public static double classSgn(double value) {
    if (value == 0.0) {
      return -1.0;
    } else {
      return 1.0;
    }
  }

  /**
    * 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 = null;

    result = new TechnicalInformation(Type.TECHREPORT);
    result.setValue(Field.AUTHOR, "Alexander Genkin and David D. Lewis and David Madigan");
    result.setValue(Field.YEAR, "2004");
    result.setValue(Field.TITLE, "Large-scale bayesian logistic regression for text categorization");
    result.setValue(Field.INSTITUTION, "DIMACS");
    result.setValue(Field.URL, "http://www.stat.rutgers.edu/~madigan/PAPERS/shortFat-v3a.pdf");
    return result;
  }

  /**
   * This is a convient function that defines and upper bound
   * (Delta>0) for values of r(i) reachable by updates in the
   * trust region.
   *
   * r BetaVector X x(i)y(i).
   * delta A parameter where sigma > 0
   * @return double function value
   */