graftsplit.java

来自「Weka」· Java 代码 · 共 541 行 · 第 1/2 页

/*
 *    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.
 */

/*
 *  GraftSplit.java
 *  Copyright (C) 2007 Geoff Webb & Janice Boughton
 *  a split object for nodes added to a tree during grafting.
 *  (used in classifier J48g).
 */

package weka.classifiers.trees.j48;

import weka.core.*;

/**
 * Class implementing a split for nodes added to a tree during grafting.
 *
 * @author Janice Boughton ([e-mail])
 * @version $Revision 1.0 $
 */
public class GraftSplit extends ClassifierSplitModel implements Comparable {

  /** the distribution for graft values, from cases in atbop */
  private Distribution m_graftdistro;
	
  /** the attribute we are splitting on */
  private int m_attIndex;

  /** value of split point (if numeric attribute) */
  private double m_splitPoint;

  /** dominant class of the subset specified by m_testType */
  private int m_maxClass;

  /** dominant class of the subset not specified by m_testType */
  private int m_otherLeafMaxClass;

  /** laplace value of the subset specified by m_testType for m_maxClass */
  private double m_laplace;

  /** leaf for the subset specified by m_testType */
  private Distribution m_leafdistro;

  /** 
   * type of test:
   * 0: <= test
   * 1: > test
   * 2: = test
   * 3: != test
   */
  private int m_testType;


  /**
   * constructor
   *
   * @param a the attribute to split on
   * @param v the value of a where split occurs
   * @param t the test type (0 is <=, 1 is >, 2 is =, 3 is !)
   * @param c the class to label the leaf node pointed to by test as.
   * @param l the laplace value (needed when sorting GraftSplits)
   */
  public GraftSplit(int a, double v, int t, double c, double l) {

    m_attIndex = a;
    m_splitPoint = v;
    m_testType = t;
    m_maxClass = (int)c;
    m_laplace = l;
  }


  /**
   * constructor
   *
   * @param a the attribute to split on
   * @param v the value of a where split occurs
   * @param t the test type (0 is <=, 1 is >, 2 is =, 3 is !=)
   * @param oC the class to label the leaf node not pointed to by test as.
   * @param counts the distribution for this split
   */
  public GraftSplit(int a, double v, int t, double oC, double [][] counts)
                                                           throws Exception {
    m_attIndex = a;
    m_splitPoint = v;
    m_testType = t;
    m_otherLeafMaxClass = (int)oC;

    // only deal with binary cuts (<= and >; = and !=)
    m_numSubsets = 2;

    // which subset are we looking at for the graft?
    int subset = subsetOfInterest();  // this is the subset for m_leaf

    // create graft distribution, based on counts
    m_distribution = new Distribution(counts);

    // create a distribution object for m_leaf
    double [][] lcounts = new double[1][m_distribution.numClasses()];
    for(int c = 0; c < lcounts[0].length; c++) {
       lcounts[0][c] = counts[subset][c];
    }
    m_leafdistro = new Distribution(lcounts);

    // set the max class
    m_maxClass = m_distribution.maxClass(subset);
 
    // set the laplace value (assumes binary class) for subset of interest
    m_laplace = (m_distribution.perClassPerBag(subset, m_maxClass) + 1.0) 
               / (m_distribution.perBag(subset) + 2.0);
  }


  /**
   * deletes the cases in data that belong to leaf pointed to by
   * the test (i.e. the subset of interest).  this is useful so
   * the instances belonging to that leaf aren't passed down the
   * other branch.
   *
   * @param data the instances to delete from
   */
  public void deleteGraftedCases(Instances data) {

    int subOfInterest = subsetOfInterest();
    for(int x = 0; x < data.numInstances(); x++) {
       if(whichSubset(data.instance(x)) == subOfInterest) {
          data.delete(x--);
       }
    }
  }


  /**
   * builds m_graftdistro using the passed data
   *
   * @param data the instances to use when creating the distribution
   */
  public void buildClassifier(Instances data) throws Exception {

    // distribution for the graft, not counting cases in atbop, only orig leaf
    m_graftdistro = new Distribution(2, data.numClasses());
 
    // which subset are we looking at for the graft?
    int subset = subsetOfInterest();  // this is the subset for m_leaf

    double thisNodeCount = 0;
    double knownCases = 0;
    boolean allKnown = true;
    // populate distribution
    for(int x = 0; x < data.numInstances(); x++) {
       Instance instance = data.instance(x);
       if(instance.isMissing(m_attIndex)) {
          allKnown = false;
          continue;
       }
       knownCases += instance.weight();
       int subst = whichSubset(instance);
       if(subst == -1)
          continue;
       m_graftdistro.add(subst, instance);
       if(subst == subset) {  // instance belongs at m_leaf
          thisNodeCount += instance.weight();
       }
    }
    double factor = (knownCases == 0) ? (1.0 / (double)2.0)
                                      : (thisNodeCount / knownCases);
    if(!allKnown) {
       for(int x = 0; x < data.numInstances(); x++) {
          if(data.instance(x).isMissing(m_attIndex)) {
             Instance instance = data.instance(x);
             int subst = whichSubset(instance);
             if(subst == -1)
                continue;
             instance.setWeight(instance.weight() * factor);
             m_graftdistro.add(subst, instance);
          }
       }
    }

    // if there are no cases at the leaf, make sure the desired
    // class is chosen, by setting counts to 0.01
    if(m_graftdistro.perBag(subset) == 0) {
       double [] counts = new double[data.numClasses()];
       counts[m_maxClass] = 0.01;
       m_graftdistro.add(subset, counts);
    }
    if(m_graftdistro.perBag((subset == 0) ? 1 : 0) == 0) {
       double [] counts = new double[data.numClasses()];
       counts[(int)m_otherLeafMaxClass] = 0.01;
       m_graftdistro.add((subset == 0) ? 1 : 0, counts);
    }
  }


  /**
   * @return the NoSplit object for the leaf pointed to by m_testType branch
   */
  public NoSplit getLeaf() {
    return new NoSplit(m_leafdistro);
  }


  /**
   * @return the NoSplit object for the leaf not pointed to by m_testType branch
   */
  public NoSplit getOtherLeaf() {

    // the bag (subset) that isn't pointed to by m_testType branch
    int bag = (subsetOfInterest() == 0) ? 1 : 0;

    double [][] counts = new double[1][m_graftdistro.numClasses()];
    double totals = 0;
    for(int c = 0; c < counts[0].length; c++) {
       counts[0][c] = m_graftdistro.perClassPerBag(bag, c);
       totals += counts[0][c];
    }
    // if empty, make sure proper class gets chosen
    if(totals == 0) {
       counts[0][m_otherLeafMaxClass] += 0.01;
    }
    return new NoSplit(new Distribution(counts));
  }


  /**
   * Prints label for subset index of instances (eg class).
   *
   * @param index the bag to dump label for
   * @param data to get attribute names and such
   * @return the label as a string
   * @exception Exception if something goes wrong
   */
  public final String dumpLabelG(int index, Instances data) throws Exception {

    StringBuffer text;

    text = new StringBuffer();
    text.append(((Instances)data).classAttribute().
       value((index==subsetOfInterest()) ? m_maxClass : m_otherLeafMaxClass));
    text.append(" ("+Utils.roundDouble(m_graftdistro.perBag(index),1));
    if(Utils.gr(m_graftdistro.numIncorrect(index),0))
       text.append("/"
        +Utils.roundDouble(m_graftdistro.numIncorrect(index),2));

    // show the graft values, only if this is subsetOfInterest()
    if(index == subsetOfInterest()) {
       text.append("|"+Utils.roundDouble(m_distribution.perBag(index),2));
       if(Utils.gr(m_distribution.numIncorrect(index),0))
          text.append("/"
             +Utils.roundDouble(m_distribution.numIncorrect(index),2));
    }
    text.append(")");
    return text.toString();
  }


  /**
   * @return the subset that is specified by the test type