Package weka.classifiers.mi

Source Code of weka.classifiers.mi.TLDSimple_Optm

/*
*    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.
*/

/*
* TLDSimple.java
* Copyright (C) 2005 University of Waikato, Hamilton, New Zealand
*
*/

package weka.classifiers.mi;

import weka.classifiers.RandomizableClassifier;
import weka.core.Capabilities;
import weka.core.Instance;
import weka.core.Instances;
import weka.core.MultiInstanceCapabilitiesHandler;
import weka.core.Optimization;
import weka.core.Option;
import weka.core.OptionHandler;
import weka.core.RevisionUtils;
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 java.util.Enumeration;
import java.util.Random;
import java.util.Vector;

/**
<!-- globalinfo-start -->
* A simpler version of TLD, mu random but sigma^2 fixed and estimated via data.<br/>
* <br/>
* For more information see:<br/>
* <br/>
* Xin Xu (2003). Statistical learning in multiple instance problem. Hamilton, NZ.
* <p/>
<!-- globalinfo-end -->
*
<!-- technical-bibtex-start -->
* BibTeX:
* <pre>
* &#64;mastersthesis{Xu2003,
*    address = {Hamilton, NZ},
*    author = {Xin Xu},
*    note = {0657.594},
*    school = {University of Waikato},
*    title = {Statistical learning in multiple instance problem},
*    year = {2003}
* }
* </pre>
* <p/>
<!-- technical-bibtex-end -->
*
<!-- options-start -->
* Valid options are: <p/>
*
* <pre> -C
*  Set whether or not use empirical
*  log-odds cut-off instead of 0</pre>
*
* <pre> -R &lt;numOfRuns&gt;
*  Set the number of multiple runs
*  needed for searching the MLE.</pre>
*
* <pre> -S &lt;num&gt;
*  Random number seed.
*  (default 1)</pre>
*
* <pre> -D
*  If set, classifier is run in debug mode and
*  may output additional info to the console</pre>
*
<!-- options-end -->
*
* @author Eibe Frank (eibe@cs.waikato.ac.nz)
* @author Xin Xu (xx5@cs.waikato.ac.nz)
* @version $Revision: 1.6 $
*/
public class TLDSimple
  extends RandomizableClassifier
  implements OptionHandler, MultiInstanceCapabilitiesHandler,
             TechnicalInformationHandler {

  /** for serialization */
  static final long serialVersionUID = 9040995947243286591L;
 
  /** The mean for each attribute of each positive exemplar */
  protected double[][] m_MeanP = null;

  /** The mean for each attribute of each negative exemplar */
  protected double[][] m_MeanN = null;

  /** The effective sum of weights of each positive exemplar in each dimension*/
  protected double[][] m_SumP = null;

  /** The effective sum of weights of each negative exemplar in each dimension*/
  protected double[][] m_SumN = null;

  /** Estimated sigma^2 in positive bags*/
  protected double[] m_SgmSqP;

  /** Estimated sigma^2 in negative bags*/
  protected double[] m_SgmSqN;

  /** The parameters to be estimated for each positive exemplar*/
  protected double[] m_ParamsP = null;

  /** The parameters to be estimated for each negative exemplar*/
  protected double[] m_ParamsN = null;

  /** The dimension of each exemplar, i.e. (numAttributes-2) */
  protected int m_Dimension = 0;

  /** The class label of each exemplar */
  protected double[] m_Class = null;

  /** The number of class labels in the data */
  protected int m_NumClasses = 2;

  /** The very small number representing zero */
  static public double ZERO = 1.0e-12;  

  protected int m_Run = 1;

  protected double m_Cutoff;

  protected boolean m_UseEmpiricalCutOff = false;   

  private double[] m_LkRatio;

  private Instances m_Attribute = null;

  /**
   * Returns a string describing this filter
   *
   * @return a description of the filter suitable for
   * displaying in the explorer/experimenter gui
   */
  public String globalInfo() {
    return
        "A simpler version of TLD, mu random but sigma^2 fixed and estimated "
      + "via data.\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.MASTERSTHESIS);
    result.setValue(Field.AUTHOR, "Xin Xu");
    result.setValue(Field.YEAR, "2003");
    result.setValue(Field.TITLE, "Statistical learning in multiple instance problem");
    result.setValue(Field.SCHOOL, "University of Waikato");
    result.setValue(Field.ADDRESS, "Hamilton, NZ");
    result.setValue(Field.NOTE, "0657.594");
   
    return result;
  }

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

    // attributes
    result.enable(Capability.NOMINAL_ATTRIBUTES);
    result.enable(Capability.RELATIONAL_ATTRIBUTES);
    result.enable(Capability.MISSING_VALUES);

    // class
    result.enable(Capability.BINARY_CLASS);
    result.enable(Capability.MISSING_CLASS_VALUES);
   
    // other
    result.enable(Capability.ONLY_MULTIINSTANCE);
   
    return result;
  }

  /**
   * Returns the capabilities of this multi-instance classifier for the
   * relational data.
   *
   * @return            the capabilities of this object
   * @see               Capabilities
   */
  public Capabilities getMultiInstanceCapabilities() {
    Capabilities result = super.getCapabilities();
   
    // attributes
    result.enable(Capability.NOMINAL_ATTRIBUTES);
    result.enable(Capability.NUMERIC_ATTRIBUTES);
    result.enable(Capability.DATE_ATTRIBUTES);
    result.enable(Capability.MISSING_VALUES);

    // class
    result.disableAllClasses();
    result.enable(Capability.NO_CLASS);
   
    return result;
  }

  /**
   *
   * @param exs the training exemplars
   * @throws Exception if the model cannot be built properly
   */   
  public void buildClassifier(Instances exs)throws Exception{
    // can classifier handle the data?
    getCapabilities().testWithFail(exs);

    // remove instances with missing class
    exs = new Instances(exs);
    exs.deleteWithMissingClass();
   
    int numegs = exs.numInstances();
    m_Dimension = exs.attribute(1).relation().numAttributes();
    m_Attribute = exs.attribute(1).relation().stringFreeStructure();
    Instances pos = new Instances(exs, 0), neg = new Instances(exs, 0);

    // Divide into two groups
    for(int u=0; u<numegs; u++){
      Instance example = exs.instance(u);
      if(example.classValue() == 1)
        pos.add(example);
      else
        neg.add(example);
   
    int pnum = pos.numInstances(), nnum = neg.numInstances()

    // xBar, n
    m_MeanP = new double[pnum][m_Dimension];
    m_SumP = new double[pnum][m_Dimension];
    m_MeanN = new double[nnum][m_Dimension];
    m_SumN = new double[nnum][m_Dimension];
    // w, m
    m_ParamsP = new double[2*m_Dimension];
    m_ParamsN = new double[2*m_Dimension];
    // \sigma^2
    m_SgmSqP = new double[m_Dimension];
    m_SgmSqN = new double[m_Dimension];
    // S^2
    double[][] varP=new double[pnum][m_Dimension],
      varN=new double[nnum][m_Dimension];
    // numOfEx 'e' without all missing
    double[] effNumExP=new double[m_Dimension],
      effNumExN=new double[m_Dimension];
    // For the starting values
    double[] pMM=new double[m_Dimension],
      nMM=new double[m_Dimension],
      pVM=new double[m_Dimension],
      nVM=new double[m_Dimension];
    // # of exemplars with only one instance
    double[] numOneInsExsP=new double[m_Dimension],
      numOneInsExsN=new double[m_Dimension];
    // sum_i(1/n_i)
    double[] pInvN = new double[m_Dimension], nInvN = new double[m_Dimension];

    // Extract metadata from both positive and negative bags
    for(int v=0; v < pnum; v++){
      //Instance px = pos.instance(v);
      Instances pxi =  pos.instance(v).relationalValue(1);
      for (int k=0; k<pxi.numAttributes(); k++) {
        m_MeanP[v][k] = pxi.meanOrMode(k);
        varP[v][k] = pxi.variance(k);
      }

      for (int w=0,t=0; w < m_Dimension; w++,t++){   
        //if((t==m_ClassIndex) || (t==m_IdIndex))
        //  t++; 
        if(varP[v][w] <= 0.0)
          varP[v][w] = 0.0;
        if(!Double.isNaN(m_MeanP[v][w])){

          for(int u=0;u<pxi.numInstances();u++)
            if(!pxi.instance(u).isMissing(t))         
              m_SumP[v][w] += pxi.instance(u).weight();

          pMM[w] += m_MeanP[v][w];
          pVM[w] += m_MeanP[v][w]*m_MeanP[v][w];       
          if((m_SumP[v][w]>1) && (varP[v][w]>ZERO)){ 

            m_SgmSqP[w] += varP[v][w]*(m_SumP[v][w]-1.0)/m_SumP[v][w];

            //m_SgmSqP[w] += varP[v][w]*(m_SumP[v][w]-1.0);
            effNumExP[w]++; // Not count exemplars with 1 instance
            pInvN[w] += 1.0/m_SumP[v][w];
            //pInvN[w] += m_SumP[v][w];
          }
          else
            numOneInsExsP[w]++;
        }

      }         
    }


    for(int v=0; v < nnum; v++){
      //Instance nx = neg.instance(v);
      Instances nxi = neg.instance(v).relationalValue(1);
      for (int k=0; k<nxi.numAttributes(); k++) {
        m_MeanN[v][k] = nxi.meanOrMode(k);
        varN[v][k] = nxi.variance(k);
      }
      //Instances nxi =  nx.getInstances();

      for (int w=0,t=0; w < m_Dimension; w++,t++){

        //if((t==m_ClassIndex) || (t==m_IdIndex))
        //  t++; 
        if(varN[v][w] <= 0.0)
          varN[v][w] = 0.0;
        if(!Double.isNaN(m_MeanN[v][w])){
          for(int u=0;u<nxi.numInstances();u++)
            if(!nxi.instance(u).isMissing(t))
              m_SumN[v][w] += nxi.instance(u).weight()

          nMM[w] += m_MeanN[v][w];
          nVM[w] += m_MeanN[v][w]*m_MeanN[v][w];
          if((m_SumN[v][w]>1) && (varN[v][w]>ZERO)){     
            m_SgmSqN[w] += varN[v][w]*(m_SumN[v][w]-1.0)/m_SumN[v][w];
            //m_SgmSqN[w] += varN[v][w]*(m_SumN[v][w]-1.0);
            effNumExN[w]++; // Not count exemplars with 1 instance
            nInvN[w] += 1.0/m_SumN[v][w];
            //nInvN[w] += m_SumN[v][w];
          }
          else
            numOneInsExsN[w]++;
        }         
      }
    }

    // Expected \sigma^2
    /* if m_SgmSqP[u] or m_SgmSqN[u] is 0, assign 0 to sigma^2.
     * Otherwise, may cause k m_SgmSqP / m_SgmSqN to be NaN.
     * Modified by Lin Dong (Sep. 2005)
     */
    for (int u=0; u < m_Dimension; u++){
      // For exemplars with only one instance, use avg(\sigma^2) of other exemplars
      if (m_SgmSqP[u]!=0)
        m_SgmSqP[u] /= (effNumExP[u]-pInvN[u]);
      else
        m_SgmSqP[u] = 0;
      if (m_SgmSqN[u]!=0)
        m_SgmSqN[u] /= (effNumExN[u]-nInvN[u]);
      else
        m_SgmSqN[u] = 0;

      //m_SgmSqP[u] /= (pInvN[u]-effNumExP[u]);
      //m_SgmSqN[u] /= (nInvN[u]-effNumExN[u]);
      effNumExP[u] += numOneInsExsP[u];
      effNumExN[u] += numOneInsExsN[u];
      pMM[u] /= effNumExP[u];
      nMM[u] /= effNumExN[u];
      pVM[u] = pVM[u]/(effNumExP[u]-1.0) - pMM[u]*pMM[u]*effNumExP[u]/(effNumExP[u]-1.0);
      nVM[u] = nVM[u]/(effNumExN[u]-1.0) - nMM[u]*nMM[u]*effNumExN[u]/(effNumExN[u]-1.0);
    }

    //Bounds and parameter values for each run
    double[][] bounds = new double[2][2];
    double[] pThisParam = new double[2],
      nThisParam = new double[2];

    // Initial values for parameters
    double w, m;
    Random whichEx = new Random(m_Seed);

    // Optimize for one dimension
    for (int x=0; x < m_Dimension; x++){    
      // System.out.println("\n\n!!!!!!!!!!!!!!!!!!!!!!???Dimension #"+x);

      // Positive examplars: first run
      pThisParam[0] = pVM[x]// w
      if( pThisParam[0] <= ZERO)
        pThisParam[0] = 1.0;
      pThisParam[1] = pMM[x]// m

      // Negative examplars: first run
      nThisParam[0] = nVM[x]// w
      if(nThisParam[0] <= ZERO)
        nThisParam[0] = 1.0;
      nThisParam[1] = nMM[x]// m

      // Bound constraints
      bounds[0][0] = ZERO; // w > 0
      bounds[0][1] = Double.NaN;
      bounds[1][0] = Double.NaN;
      bounds[1][1] = Double.NaN;

      double pminVal=Double.MAX_VALUE, nminVal=Double.MAX_VALUE;
      TLDSimple_Optm pOp=null, nOp=null
      boolean isRunValid = true;
      double[] sumP=new double[pnum], meanP=new double[pnum];
      double[] sumN=new double[nnum], meanN=new double[nnum];

      // One dimension
      for(int p=0; p<pnum; p++){
        sumP[p] = m_SumP[p][x];
        meanP[p] = m_MeanP[p][x];
      }
      for(int q=0; q<nnum; q++){
        sumN[q] = m_SumN[q][x];
        meanN[q] = m_MeanN[q][x];
      }

      for(int y=0; y<m_Run; y++){
        //System.out.println("\n\n!!!!!!!!!Positive exemplars: Run #"+y);
        double thisMin;
        pOp = new TLDSimple_Optm();
        pOp.setNum(sumP);
        pOp.setSgmSq(m_SgmSqP[x]);
        if (getDebug())
          System.out.println("m_SgmSqP["+x+"]= " +m_SgmSqP[x]);
        pOp.setXBar(meanP);
        //pOp.setDebug(true);
        pThisParam = pOp.findArgmin(pThisParam, bounds);
        while(pThisParam==null){
          pThisParam = pOp.getVarbValues();       
          if (getDebug())
            System.out.println("!!! 200 iterations finished, not enough!");
          pThisParam = pOp.findArgmin(pThisParam, bounds);
       

        thisMin = pOp.getMinFunction();
        if(!Double.isNaN(thisMin) && (thisMin<pminVal)){
          pminVal = thisMin;
          for(int z=0; z<2; z++)
            m_ParamsP[2*x+z] = pThisParam[z];
        }

        if(Double.isNaN(thisMin)){
          pThisParam = new double[2];
          isRunValid =false;
        }
        if(!isRunValid){ y--; isRunValid=true; }

        // Change the initial parameters and restart
        int pone = whichEx.nextInt(pnum);

        // Positive exemplars: next run
        while(Double.isNaN(m_MeanP[pone][x]))
          pone = whichEx.nextInt(pnum);

        m = m_MeanP[pone][x];
        w = (m-pThisParam[1])*(m-pThisParam[1]);
        pThisParam[0] = w;  // w
        pThisParam[1] = m;  // m     
      }

      for(int y=0; y<m_Run; y++){
        //System.out.println("\n\n!!!!!!!!!Negative exemplars: Run #"+y);
        double thisMin;
        nOp = new TLDSimple_Optm();
        nOp.setNum(sumN);
        nOp.setSgmSq(m_SgmSqN[x]);
        if (getDebug())
          System.out.println(m_SgmSqN[x]);
        nOp.setXBar(meanN);
        //nOp.setDebug(true);
        nThisParam = nOp.findArgmin(nThisParam, bounds);

        while(nThisParam==null){ 
          nThisParam = nOp.getVarbValues();
          if (getDebug())
            System.out.println("!!! 200 iterations finished, not enough!");
          nThisParam = nOp.findArgmin(nThisParam, bounds);
        }     

        thisMin = nOp.getMinFunction()
        if(!Double.isNaN(thisMin) && (thisMin<nminVal)){
          nminVal = thisMin;
          for(int z=0; z<2; z++)
            m_ParamsN[2*x+z] = nThisParam[z];    
        }

        if(Double.isNaN(thisMin)){
          nThisParam = new double[2];
          isRunValid =false;
        }

        if(!isRunValid){ y--; isRunValid=true; }    

        // Change the initial parameters and restart          
        int none = whichEx.nextInt(nnum);// Randomly pick one pos. exmpl.

        // Negative exemplars: next run
        while(Double.isNaN(m_MeanN[none][x]))
          none = whichEx.nextInt(nnum);

        m = m_MeanN[none][x];
        w = (m-nThisParam[1])*(m-nThisParam[1]);
        nThisParam[0] = w;  // w
        nThisParam[1] = m;  // m      
      }                 
    }

    m_LkRatio = new double[m_Dimension];

    if(m_UseEmpiricalCutOff){ 
      // Find the empirical cut-off
      double[] pLogOdds=new double[pnum], nLogOdds=new double[nnum]
      for(int p=0; p<pnum; p++)
        pLogOdds[p] =
          likelihoodRatio(m_SumP[p], m_MeanP[p]);

      for(int q=0; q<nnum; q++)
        nLogOdds[q] =
          likelihoodRatio(m_SumN[q], m_MeanN[q]);

      // Update m_Cutoff
      findCutOff(pLogOdds, nLogOdds);
    }
    else
      m_Cutoff = -Math.log((double)pnum/(double)nnum);

    /*
       for(int x=0, y=0; x<m_Dimension; x++, y++){
       if((x==exs.classIndex()) || (x==exs.idIndex()))
       y++;

       w=m_ParamsP[2*x]; m=m_ParamsP[2*x+1];
       System.err.println("\n\n???Positive: ( "+exs.attribute(y)+
       "):  w="+w+", m="+m+", sgmSq="+m_SgmSqP[x]);

       w=m_ParamsN[2*x]; m=m_ParamsN[2*x+1];
       System.err.println("???Negative: ("+exs.attribute(y)+
       "):  w="+w+", m="+m+", sgmSq="+m_SgmSqN[x]+
       "\nAvg. log-likelihood ratio in training data="
       +(m_LkRatio[x]/(pnum+nnum)));
       } 
       */
    if (getDebug())
      System.err.println("\n\n???Cut-off="+m_Cutoff);
  }       

  /**
   *
   * @param ex the given test exemplar
   * @return the classification
   * @throws Exception if the exemplar could not be classified
   * successfully
   */
  public double classifyInstance(Instance ex)throws Exception{
    //Instance ex = new Exemplar(e);
    Instances exi = ex.relationalValue(1);
    double[] n = new double[m_Dimension];
    double [] xBar = new double[m_Dimension];
    for (int i=0; i<exi.numAttributes() ; i++)
      xBar[i] = exi.meanOrMode(i);

    for (int w=0, t=0; w < m_Dimension; w++, t++){
      // if((t==m_ClassIndex) || (t==m_IdIndex))
      //t++; 
      for(int u=0;u<exi.numInstances();u++)
        if(!exi.instance(u).isMissing(t))
          n[w] += exi.instance(u).weight();
    }

    double logOdds = likelihoodRatio(n, xBar);
    return (logOdds > m_Cutoff) ? 1 : 0 ;
  }
 
  /**
   * Computes the distribution for a given exemplar
   *
   * @param ex the exemplar for which distribution is computed
   * @return the distribution
   * @throws Exception if the distribution can't be computed successfully
   */
  public double[] distributionForInstance(Instance ex) throws Exception {
   
    double[] distribution = new double[2];
    Instances exi = ex.relationalValue(1);
    double[] n = new double[m_Dimension];
    double[] xBar = new double[m_Dimension];
    for (int i = 0; i < exi.numAttributes() ; i++)
      xBar[i] = exi.meanOrMode(i);
   
    for (int w = 0, t = 0; w < m_Dimension; w++, t++){
      for (int u = 0; u < exi.numInstances(); u++)
  if (!exi.instance(u).isMissing(t))
    n[w] += exi.instance(u).weight();
    }
   
    double logOdds = likelihoodRatio(n, xBar);
   
    // returned logOdds value has been divided by m_Dimension to avoid
    // Math.exp(logOdds) getting too large or too small,
    // that may result in two fixed distribution value (1 or 0).
    distribution[0] = 1 / (1 + Math.exp(logOdds)); // Prob. for class 0 (negative)
    distribution[1] = 1 - distribution[0];
   
    return distribution;
 

  /**
   * Compute the log-likelihood ratio
   */
  private double likelihoodRatio(double[] n, double[] xBar){ 
    double LLP = 0.0, LLN = 0.0;

    for (int x=0; x<m_Dimension; x++){
      if(Double.isNaN(xBar[x])) continue; // All missing values
      //if(Double.isNaN(xBar[x]) || (m_ParamsP[2*x] <= ZERO)
      //  || (m_ParamsN[2*x]<=ZERO))
      //  continue; // All missing values

      //Log-likelihood for positive
      double w=m_ParamsP[2*x], m=m_ParamsP[2*x+1];
      double llp = Math.log(w*n[x]+m_SgmSqP[x])
        + n[x]*(m-xBar[x])*(m-xBar[x])/(w*n[x]+m_SgmSqP[x]);
      LLP -= llp;

      //Log-likelihood for negative
      w=m_ParamsN[2*x]; m=m_ParamsN[2*x+1];
      double lln = Math.log(w*n[x]+m_SgmSqN[x])
        + n[x]*(m-xBar[x])*(m-xBar[x])/(w*n[x]+m_SgmSqN[x]);
      LLN -= lln;

      m_LkRatio[x] += llp - lln;
    }

    return LLP - LLN / m_Dimension;
  }

  private void findCutOff(double[] pos, double[] neg){
    int[] pOrder = Utils.sort(pos),
      nOrder = Utils.sort(neg);
    /*
       System.err.println("\n\n???Positive: ");
       for(int t=0; t<pOrder.length; t++)
       System.err.print(t+":"+Utils.doubleToString(pos[pOrder[t]],0,2)+" ");
       System.err.println("\n\n???Negative: ");
       for(int t=0; t<nOrder.length; t++)
       System.err.print(t+":"+Utils.doubleToString(neg[nOrder[t]],0,2)+" ");
       */
    int pNum = pos.length, nNum = neg.length, count, p=0, n=0
    double fstAccu=0.0, sndAccu=(double)pNum, split;
    double maxAccu = 0, minDistTo0 = Double.MAX_VALUE;

    // Skip continuous negatives 
    for(;(n<nNum)&&(pos[pOrder[0]]>=neg[nOrder[n]]); n++, fstAccu++);

    if(n>=nNum){ // totally seperate
      m_Cutoff = (neg[nOrder[nNum-1]]+pos[pOrder[0]])/2.0
      //m_Cutoff = neg[nOrder[nNum-1]];
      return
   

    count=n;
    while((p<pNum)&&(n<nNum)){
      // Compare the next in the two lists
      if(pos[pOrder[p]]>=neg[nOrder[n]]){ // Neg has less log-odds
        fstAccu += 1.0;   
        split=neg[nOrder[n]];
        n++;  
      }
      else{
        sndAccu -= 1.0;
        split=pos[pOrder[p]];
        p++;
      }         
      count++;
      /*
         double entropy=0.0, cover=(double)count;
         if(fstAccu>0.0)
         entropy -= fstAccu*Math.log(fstAccu/cover);
         if(sndAccu>0.0)
         entropy -= sndAccu*Math.log(sndAccu/(total-cover));

         if(entropy < minEntropy){
         minEntropy = entropy;
      //find the next smallest
      //double next = neg[nOrder[n]];
      //if(pos[pOrder[p]]<neg[nOrder[n]])
      //    next = pos[pOrder[p]]; 
      //m_Cutoff = (split+next)/2.0;
      m_Cutoff = split;
         }
         */
      if ((fstAccu+sndAccu > maxAccu) ||
          ((fstAccu+sndAccu == maxAccu) && (Math.abs(split)<minDistTo0))){
        maxAccu = fstAccu+sndAccu;
        m_Cutoff = split;
        minDistTo0 = Math.abs(split);
     }     
    }   
  }

  /**
   * Returns an enumeration describing the available options
   *
   * @return an enumeration of all the available options
   */
  public Enumeration listOptions() {
    Vector result = new Vector();
   
    result.addElement(new Option(
          "\tSet whether or not use empirical\n"
          + "\tlog-odds cut-off instead of 0",
          "C", 0, "-C"));
   
    result.addElement(new Option(
          "\tSet the number of multiple runs \n"
          + "\tneeded for searching the MLE.",
          "R", 1, "-R <numOfRuns>"));
   
    Enumeration enu = super.listOptions();
    while (enu.hasMoreElements()) {
      result.addElement(enu.nextElement());
    }

    return result.elements();
  }

  /**
   * Parses a given list of options. <p/>
   *
   <!-- options-start -->
   * Valid options are: <p/>
   *
   * <pre> -C
   *  Set whether or not use empirical
   *  log-odds cut-off instead of 0</pre>
   *
   * <pre> -R &lt;numOfRuns&gt;
   *  Set the number of multiple runs
   *  needed for searching the MLE.</pre>
   *
   * <pre> -S &lt;num&gt;
   *  Random number seed.
   *  (default 1)</pre>
   *
   * <pre> -D
   *  If set, classifier is run in debug mode and
   *  may output additional info to the console</pre>
   *
   <!-- options-end -->
   *
   * @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{
    setDebug(Utils.getFlag('D', options));

    setUsingCutOff(Utils.getFlag('C', options));

    String runString = Utils.getOption('R', options);
    if (runString.length() != 0)
      setNumRuns(Integer.parseInt(runString));
    else
      setNumRuns(1);

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

    if (getDebug())
      result.add("-D");
   
    if (getUsingCutOff())
      result.add("-C");

    result.add("-R");
    result.add("" + getNumRuns());

    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 numRunsTipText() {
    return "The number of runs to perform.";
  }
 
  /**
   * Sets the number of runs to perform.
   *
   * @param numRuns   the number of runs to perform
   */
  public void setNumRuns(int numRuns) {
    m_Run = numRuns;
  }

  /**
   * Returns the number of runs to perform.
   *
   * @return          the number of runs to perform
   */
  public int getNumRuns() {
    return m_Run;
  }

  /**
   * Returns the tip text for this property
   *
   * @return tip text for this property suitable for
   * displaying in the explorer/experimenter gui
   */
  public String usingCutOffTipText() {
    return "Whether to use an empirical cutoff.";
  }

  /**
   * Sets whether to use an empirical cutoff.
   *
   * @param cutOff      whether to use an empirical cutoff
   */
  public void setUsingCutOff (boolean cutOff) {
    m_UseEmpiricalCutOff =cutOff;
  }

  /**
   * Returns whether an empirical cutoff is used
   *
   * @return            true if an empirical cutoff is used
   */
  public boolean getUsingCutOff() {
    return m_UseEmpiricalCutOff ;
  }

  /**
   * Gets a string describing the classifier.
   *
   * @return a string describing the classifer built.
   */
  public String toString(){
    StringBuffer text = new StringBuffer("\n\nTLDSimple:\n");
    double sgm, w, m;
    for (int x=0, y=0; x<m_Dimension; x++, y++){
      // if((x==m_ClassIndex) || (x==m_IdIndex))
      //y++;
      sgm = m_SgmSqP[x];
      w=m_ParamsP[2*x];
      m=m_ParamsP[2*x+1];
      text.append("\n"+m_Attribute.attribute(y).name()+"\nPositive: "+
          "sigma^2="+sgm+", w="+w+", m="+m+"\n");
      sgm = m_SgmSqN[x];
      w=m_ParamsN[2*x];
      m=m_ParamsN[2*x+1];
      text.append("Negative: "+
          "sigma^2="+sgm+", w="+w+", m="+m+"\n");
    }

    return text.toString();
  }    
 
  /**
   * Returns the revision string.
   *
   * @return    the revision
   */
  public String getRevision() {
    return RevisionUtils.extract("$Revision: 1.6 $");
  }

  /**
   * Main method for testing.
   *
   * @param args the options for the classifier
   */
  public static void main(String[] args) { 
    runClassifier(new TLDSimple(), args);
  }
}

class TLDSimple_Optm extends Optimization {

  private double[] num;
  private double sSq;
  private double[] xBar;

  public void setNum(double[] n) {num = n;}
  public void setSgmSq(double s){

    sSq = s;
  }
  public void setXBar(double[] x){xBar = x;}

  /**
   * Implement this procedure to evaluate objective
   * function to be minimized
   */
  protected double objectiveFunction(double[] x){
    int numExs = num.length;
    double NLL=0; // Negative Log-Likelihood

    double w=x[0], m=x[1];
    for(int j=0; j < numExs; j++){

      if(Double.isNaN(xBar[j])) continue; // All missing values
      double bag=0;

      bag += Math.log(w*num[j]+sSq);

      if(Double.isNaN(bag) && m_Debug){
        System.out.println("???????????1: "+w+" "+m
            +"|x-: "+xBar[j] +
            "|n: "+num[j] + "|S^2: "+sSq);
        //System.exit(1);
      }

      bag += num[j]*(m-xBar[j])*(m-xBar[j])/(w*num[j]+sSq);           
      if(Double.isNaN(bag) && m_Debug){
        System.out.println("???????????2: "+w+" "+m
            +"|x-: "+xBar[j] +
            "|n: "+num[j] + "|S^2: "+sSq);
        //System.exit(1);
      }              

      //if(bag<0) bag=0;
      NLL += bag;
    }

    //System.out.println("???????????NLL:"+NLL);
    return NLL;
  }

  /**
   * Subclass should implement this procedure to evaluate gradient
   * of the objective function
   */
  protected double[] evaluateGradient(double[] x){
    double[] g = new double[x.length];
    int numExs = num.length;

    double w=x[0],m=x[1]
    double dw=0.0, dm=0.0;

    for(int j=0; j < numExs; j++){

      if(Double.isNaN(xBar[j])) continue; // All missing values     
      dw += num[j]/(w*num[j]+sSq)
        - num[j]*num[j]*(m-xBar[j])*(m-xBar[j])/((w*num[j]+sSq)*(w*num[j]+sSq));

      dm += 2.0*num[j]*(m-xBar[j])/(w*num[j]+sSq);
    }

    g[0] = dw;
    g[1] = dm;
    return g;
  }

  /**
   * Subclass should implement this procedure to evaluate second-order
   * gradient of the objective function
   */
  protected double[] evaluateHessian(double[] x, int index){
    double[] h = new double[x.length];

    // # of exemplars, # of dimensions
    // which dimension and which variable for 'index'
    int numExs = num.length;
    double w,m;
    // Take the 2nd-order derivative
    switch(index){ 
      case 0: // w  
        w=x[0];m=x[1];

        for(int j=0; j < numExs; j++){
          if(Double.isNaN(xBar[j])) continue; //All missing values

          h[0] += 2.0*Math.pow(num[j],3)*(m-xBar[j])*(m-xBar[j])/Math.pow(w*num[j]+sSq,3)
            - num[j]*num[j]/((w*num[j]+sSq)*(w*num[j]+sSq));

          h[1] -= 2.0*(m-xBar[j])*num[j]*num[j]/((num[j]*w+sSq)*(num[j]*w+sSq));   
        }
        break;

      case 1: // m
        w=x[0];m=x[1];

        for(int j=0; j < numExs; j++){
          if(Double.isNaN(xBar[j])) continue; //All missing values

          h[0] -= 2.0*(m-xBar[j])*num[j]*num[j]/((num[j]*w+sSq)*(num[j]*w+sSq));

          h[1] += 2.0*num[j]/(w*num[j]+sSq);       
        }
    }

    return h;
  }
 
  /**
   * Returns the revision string.
   *
   * @return    the revision
   */
  public String getRevision() {
    return RevisionUtils.extract("$Revision: 1.6 $");
  }
}
TOP

Related Classes of weka.classifiers.mi.TLDSimple_Optm

TOP
Copyright © 2018 www.massapi.com. All rights reserved.
All source code are property of their respective owners. Java is a trademark of Sun Microsystems, Inc and owned by ORACLE Inc. Contact coftware#gmail.com.