removemisclassified.java

wekaUT是 university texas austin 开发的基于weka的半指导学习(semi supervised learning)的分类器

/*
 *    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.
 */

/*
 *    RemoveMisclassified.java
 *    Copyright (C) 2002 Richard Kirkby
 *
 */

package weka.filters.unsupervised.instance;

import weka.filters.*;
import weka.classifiers.Classifier;
import weka.core.*;
import java.util.Enumeration;
import java.util.Vector;

/** 
 * A filter that removes instances which are incorrectly classified.
 * Useful for removing outliers. <p>
 *
 * Valid filter-specific options are: <p>
 *
 * -W classifier string <br>
 * Full class name of classifier to use, followed by scheme options. (required)<p>
 * 
 * -C class index <br>
 * Attribute on which misclassifications are based. If < 0 will use any current set
 * class or default to the last attribute.
 *
 * -F number of folds <br>
 * The number of folds to use for cross-validation cleansing.
 * (<2 = no cross-validation - default)<p> 
 *
 * -T threshold <br>
 * Threshold for the max error when predicting numeric class.
 * (Value should be >= 0, default = 0.1)<p>
 *
 * -I max iterations <br>
 * The maximum number of cleansing iterations to perform.
 * (<1 = until fully cleansed - default)<p>
 *
 * -V <br>
 * Invert the match so that correctly classified instances are discarded.<p>
 *
 * @author Richard Kirkby ([e-mail])
 * @author Malcolm Ware ([e-mail])
 * @version $Revision: 1.1.1.1 $
 */
public class RemoveMisclassified extends Filter implements UnsupervisedFilter, OptionHandler {

  /** The classifier used to do the cleansing */
  protected Classifier m_cleansingClassifier = new weka.classifiers.rules.ZeroR();

  /** The attribute to treat as the class for purposes of cleansing. */
  protected int m_classIndex = -1;

  /** The number of cross validation folds to perform (<2 = no cross validation)  */
  protected int m_numOfCrossValidationFolds = 0;
  
  /** The maximum number of cleansing iterations to perform (<1 = until fully cleansed)  */
  protected int m_numOfCleansingIterations = 0;

  /** The threshold for deciding when a numeric value is correctly classified */
  protected double m_numericClassifyThreshold = 0.1;

  /** Whether to invert the match so the correctly classified instances are discarded */
  protected boolean m_invertMatching = false;

  /**
   * Sets the format of the input instances.
   *
   * @param instanceInfo an Instances object containing the input instance
   * structure (any instances contained in the object are ignored - only the
   * structure is required).
   * @return true if the outputFormat may be collected immediately
   * @exception Exception if the inputFormat can't be set successfully 
   */ 
  public boolean setInputFormat(Instances instanceInfo) throws Exception {
    
    super.setInputFormat(instanceInfo);
    setOutputFormat(instanceInfo);
    return true;
  }

  /**
   * Cleanses the data based on misclassifications when used training data.
   *
   * @param data the data to train with and cleanse
   */
  private Instances cleanseTrain(Instances data) throws Exception {
    
    Instance inst;
    Instances buildSet = new Instances(data);  
    Instances temp = new Instances(data, data.numInstances());
    Instances inverseSet = new Instances(data, data.numInstances()); 
    int count = 0;
    double ans;
    int iterations = 0;
    int classIndex = m_classIndex;
    if (classIndex < 0) classIndex = data.classIndex();
    if (classIndex < 0) classIndex = data.numAttributes()-1;

    // loop until perfect
    while(count != buildSet.numInstances()) {
      
      // check if hit maximum number of iterations
      iterations++;
      if (m_numOfCleansingIterations > 0 && iterations > m_numOfCleansingIterations) break;

      // build classifier
      count = buildSet.numInstances();
      buildSet.setClassIndex(classIndex);
      m_cleansingClassifier.buildClassifier(buildSet);

      temp = new Instances(buildSet, buildSet.numInstances());

      // test on training data
      for (int i = 0; i < buildSet.numInstances(); i++) {
	inst = buildSet.instance(i);
	ans = m_cleansingClassifier.classifyInstance(inst);
	if (buildSet.classAttribute().isNumeric()) {
	  if (ans >= inst.classValue() - m_numericClassifyThreshold &&
	      ans <= inst.classValue() + m_numericClassifyThreshold) {
	    temp.add(inst);
	  } else if (m_invertMatching) {
	    inverseSet.add(inst);
	  }
	}
	else { //class is nominal
	  if (ans == inst.classValue()) {
	    temp.add(inst);
	  } else if (m_invertMatching) {
	    inverseSet.add(inst);
	  }
	}
      }
      buildSet = temp;
    }

    if (m_invertMatching) {
      inverseSet.setClassIndex(data.classIndex());
      return inverseSet;
    }
    else {
      buildSet.setClassIndex(data.classIndex());
      return buildSet;
    }
  }

  /**
   * Cleanses the data based on misclassifications when performing cross-validation.
   *
   * @param data the data to train with and cleanse
   */
  private Instances cleanseCross(Instances data) throws Exception {
    
    Instance inst;
    Instances crossSet = new Instances(data);
    Instances temp = new Instances(data, data.numInstances());    
    Instances inverseSet = new Instances(data, data.numInstances()); 
    int count = 0;
    double ans;
    int iterations = 0;
    int classIndex = m_classIndex;
    if (classIndex < 0) classIndex = data.classIndex();
    if (classIndex < 0) classIndex = data.numAttributes()-1;

    // loop until perfect
    while (count != crossSet.numInstances() && 
	   crossSet.numInstances() >= m_numOfCrossValidationFolds) {

      count = crossSet.numInstances();
      
      // check if hit maximum number of iterations
      iterations++;
      if (m_numOfCleansingIterations > 0 && iterations > m_numOfCleansingIterations) break;

      crossSet.setClassIndex(classIndex);

      if (crossSet.classAttribute().isNominal()) {
	crossSet.stratify(m_numOfCrossValidationFolds);
      }
      // do the folds
      temp = new Instances(crossSet, crossSet.numInstances());
      
      for (int fold = 0; fold < m_numOfCrossValidationFolds; fold++) {
	Instances train = crossSet.trainCV(m_numOfCrossValidationFolds, fold);
	m_cleansingClassifier.buildClassifier(train);
	Instances test = crossSet.testCV(m_numOfCrossValidationFolds, fold);
	//now test
	for (int i = 0; i < test.numInstances(); i++) {
	  inst = test.instance(i);
	  ans = m_cleansingClassifier.classifyInstance(inst);
	  if (crossSet.classAttribute().isNumeric()) {
	    if (ans >= inst.classValue() - m_numericClassifyThreshold &&
		ans <= inst.classValue() + m_numericClassifyThreshold) {
	      temp.add(inst);
	    } else if (m_invertMatching) {
	      inverseSet.add(inst);
	    }
	  }
	  else { //class is nominal
	    if (ans == inst.classValue()) {
	      temp.add(inst);
	    } else if (m_invertMatching) {
	      inverseSet.add(inst);
	    }
	  }
	}
      }
      crossSet = temp;
    }

    if (m_invertMatching) {
      inverseSet.setClassIndex(data.classIndex());
      return inverseSet;
    }
    else {
      crossSet.setClassIndex(data.classIndex());
      return crossSet;
    }

  }
  
  /**
   * Signify that this batch of input to the filter is finished.
   *
   * @return true if there are instances pending output
   * @exception IllegalStateException if no input structure has been defined 
   */  
  public boolean batchFinished() throws Exception {

    if (getInputFormat() == null) {
      throw new IllegalStateException("No input instance format defined");
    }

    Instances filtered;

    if (m_numOfCrossValidationFolds < 2) {
      filtered = cleanseTrain(getInputFormat());
    } else {
      filtered = cleanseCross(getInputFormat());
    }

    for (int i=0; i<filtered.numInstances(); i++) {
      push(filtered.instance(i));
    }

    return (numPendingOutput() != 0);
  }

  /**
   * Returns an enumeration describing the available options.
   *
   * @return an enumeration of all the available options.
   */
  public Enumeration listOptions() {
    
    Vector newVector = new Vector(6);
    
    newVector.addElement(new Option(
	      "\tFull class name of classifier to use, followed\n"
	      + "\tby scheme options. (required)\n"
	      + "\teg: \"weka.classifiers.bayes.NaiveBayes -D\"",
	      "W", 1, "-W <classifier specification>"));
    newVector.addElement(new Option(
	      "\tAttribute on which misclassifications are based.\n"
	      + "\tIf < 0 will use any current set class or default to the last attribute.",
	      "C", 1, "-C <class index>"));
    newVector.addElement(new Option(
	      "\tThe number of folds to use for cross-validation cleansing.\n"
	      +"\t(<2 = no cross-validation - default).",
	      "F", 1, "-F <number of folds>"));
    newVector.addElement(new Option(
	      "\tThreshold for the max error when predicting numeric class.\n"
	      +"\t(Value should be >= 0, default = 0.1).",
	      "T", 1, "-T <threshold>"));
    newVector.addElement(new Option(
	      "\tThe maximum number of cleansing iterations to perform.\n"
	      +"\t(<1 = until fully cleansed - default)",
	      "I", 1,"-I"));
    newVector.addElement(new Option(
	      "\tInvert the match so that correctly classified instances are discarded.\n",
	      "V", 0,"-V"));
    
    return newVector.elements();
  }


  /**
   * Parses the options for this object. Valid options are: <p>
   *
   * -W classifier string <br>
   * Full class name of classifier to use, followed by scheme options. (required)<p>
   * 
   * -C class index <br>
   * Attribute on which misclassifications are based. If < 0 will use any current
   * set class or default to the last attribute.
   *
   * -F number of folds <br>
   * The number of folds to use for cross-validation cleansing.
   * (<2 = no cross-validation - default)<p> 
   *
   * -T threshold <br>
   * Threshold for the max error when predicting numeric class.
   * (Value should be >= 0, default = 0.1)<p>
   *
   * -I max iterations <br>
   * The maximum number of cleansing iterations to perform.