miemdd.java

代码是一个分类器的实现,其中使用了部分weka的源代码。可以将项目导入eclipse运行

    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;
  }

  /**
   * Builds the classifier
   *
   * @param train 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 train) throws Exception {
    // can classifier handle the data?
    getCapabilities().testWithFail(train);

    // remove instances with missing class
    train = new Instances(train);
    train.deleteWithMissingClass();
    
    m_ClassIndex = train.classIndex();
    m_NumClasses = train.numClasses();

    int nR = train.attribute(1).relation().numAttributes();
    int nC = train.numInstances();
    int[] bagSize = new int[nC];
    Instances datasets = new Instances(train.attribute(1).relation(), 0);

    m_Data = new double [nC][nR][];              // Data values
    m_Classes = new int [nC];                    // Class values
    m_Attributes = datasets.stringFreeStructure();
    if (m_Debug) {
      System.out.println("\n\nExtracting data...");
    }

    for (int h = 0; h < nC; h++)  {//h_th bag
      Instance current = train.instance(h);
      m_Classes[h] = (int)current.classValue();  // Class value starts from 0
      Instances currInsts = current.relationalValue(1);
      for (int i = 0; i < currInsts.numInstances(); i++){
        Instance inst = currInsts.instance(i);
        datasets.add(inst);
      }

      int nI = currInsts.numInstances();
      bagSize[h] = nI;
    }


    /* filter the training data */
    if (m_filterType == FILTER_STANDARDIZE)  
      m_Filter = new Standardize();
    else if (m_filterType == FILTER_NORMALIZE)
      m_Filter = new Normalize();
    else 
      m_Filter = null; 

    if (m_Filter != null) {    
      m_Filter.setInputFormat(datasets);
      datasets = Filter.useFilter(datasets, m_Filter); 	
    }

    m_Missing.setInputFormat(datasets);
    datasets = Filter.useFilter(datasets, m_Missing);

    int instIndex = 0;
    int start = 0;	
    for (int h = 0; h < nC; h++)  {	
      for (int i = 0; i < datasets.numAttributes(); i++) {
        // initialize m_data[][][]
        m_Data[h][i] = new double[bagSize[h]];
        instIndex=start;
        for (int k = 0; k < bagSize[h]; k++){
          m_Data[h][i][k] = datasets.instance(instIndex).value(i);
          instIndex++;
        }
      }
      start=instIndex;
    }

    if (m_Debug) {
      System.out.println("\n\nIteration History..." );
    }

    m_emData =new double[nC][nR];
    m_Par= new double[2*nR];

    double[] x = new double[nR*2];
    double[] tmp = new double[x.length];
    double[] pre_x = new double[x.length];
    double[] best_hypothesis = new double[x.length];
    double[][] b = new double[2][x.length];

    OptEng opt;
    double bestnll = Double.MAX_VALUE;
    double min_error = Double.MAX_VALUE;
    double nll, pre_nll;
    int iterationCount;


    for (int t = 0; t < x.length; t++) {
      b[0][t] = Double.NaN;
      b[1][t] = Double.NaN;
    }

    //random pick 3 positive bags 
    Random r = new Random(getSeed());
    FastVector index = new FastVector(); 
    int n1, n2, n3;
    do {
      n1 = r.nextInt(nC-1);	
    } while (m_Classes[n1] == 0);
    index.addElement(new Integer(n1)); 

    do {
      n2 = r.nextInt(nC-1);
    } while (n2 == n1|| m_Classes[n2] == 0);
    index.addElement(new Integer(n2)); 

    do {
      n3 = r.nextInt(nC-1);
    } while (n3 == n1 || n3 == n2 || m_Classes[n3] == 0);
    index.addElement(new Integer(n3));

    for (int s = 0; s < index.size(); s++){
      int exIdx = ((Integer)index.elementAt(s)).intValue();
      if (m_Debug)
        System.out.println("\nH0 at "+exIdx);


      for (int p = 0; p < m_Data[exIdx][0].length; p++) {
        //initialize a hypothesis
        for (int q = 0; q < nR; q++) {
          x[2 * q] = m_Data[exIdx][q][p];
          x[2 * q + 1] = 1.0;
        } 

        pre_nll = Double.MAX_VALUE;
        nll = Double.MAX_VALUE/10.0;
        iterationCount = 0;
        //while (Math.abs(nll-pre_nll)>0.01*pre_nll && iterationCount<10) {  //stop condition
        while (nll < pre_nll && iterationCount < 10) {
          iterationCount++;
          pre_nll = nll;

          if (m_Debug) 
            System.out.println("\niteration: "+iterationCount);

          //E-step (find one instance from each bag with max likelihood )
          for (int i = 0; i < m_Data.length; i++) { //for each bag

            int insIndex = findInstance(i, x); 

            for (int att = 0; att < m_Data[0].length; att++) //for each attribute
              m_emData[i][att] = m_Data[i][att][insIndex];
          }
          if (m_Debug)
            System.out.println("E-step for new H' finished");

          //M-step
          opt = new OptEng();
          tmp = opt.findArgmin(x, b);
          while (tmp == null) {
            tmp = opt.getVarbValues();
            if (m_Debug)
              System.out.println("200 iterations finished, not enough!");
            tmp = opt.findArgmin(tmp, b);
          }
          nll = opt.getMinFunction();

          pre_x = x;
          x = tmp; // update hypothesis 


          //keep the track of the best target point which has the minimum nll
          /* if (nll < bestnll) {
             bestnll = nll;
             m_Par = tmp;
             if (m_Debug)
             System.out.println("!!!!!!!!!!!!!!!!Smaller NLL found: " + nll);
             }*/

          //if (m_Debug)
          //System.out.println(exIdx+" "+p+": "+nll+" "+pre_nll+" " +bestnll);

        } //converged for one instance

        //evaluate the hypothesis on the training data and
        //keep the track of the hypothesis with minimum error on training data
        double distribution[] = new double[2];
        int error = 0;
        if (nll > pre_nll)
          m_Par = pre_x; 
        else
          m_Par = x;

        for (int i = 0; i<train.numInstances(); i++) {
          distribution = distributionForInstance (train.instance(i));
          if (distribution[1] >= 0.5 && m_Classes[i] == 0)
            error++;
          else if (distribution[1]<0.5 && m_Classes[i] == 1)
            error++;
        }
        if (error < min_error) {
          best_hypothesis = m_Par;
          min_error = error;
          if (nll > pre_nll)
            bestnll = pre_nll;
          else
            bestnll = nll;
          if (m_Debug)
            System.out.println("error= "+ error +"  nll= " + bestnll);
        }
      }
      if (m_Debug) {
        System.out.println(exIdx+ ":  -------------<Converged>--------------");
        System.out.println("current minimum error= "+min_error+"  nll= "+bestnll);
      }
    } 
    m_Par = best_hypothesis;
  }


  /**
   * given x, find the instance in ith bag with the most likelihood
   * probability, which is most likely to responsible for the label of the
   * bag For a positive bag, find the instance with the maximal probability
   * of being positive For a negative bag, find the instance with the minimal
   * probability of being negative
   *
   * @param i the bag index
   * @param x the current values of variables
   * @return index of the instance in the bag
   */
  protected int findInstance(int i, double[] x){

    double min=Double.MAX_VALUE;
    int insIndex=0;
    int nI = m_Data[i][0].length; // numInstances in ith bag

    for (int j=0; j<nI; j++){
      double ins=0.0;
      for (int k=0; k<m_Data[i].length; k++)  // for each attribute
        ins += (m_Data[i][k][j]-x[k*2])*(m_Data[i][k][j]-x[k*2])*
          x[k*2+1]*x[k*2+1];

      //the probability can be calculated as Math.exp(-ins)
      //to find the maximum Math.exp(-ins) is equivalent to find the minimum of (ins)
      if (ins<min)  {
        min=ins;
        insIndex=j;
      }
    }
    return insIndex;
  }


  /**
   * Computes the distribution for a given exemplar
   *
   * @param exmp the exemplar for which distribution is computed
   * @return the distribution
   * @throws Exception if the distribution can't be computed successfully
   */
  public double[] distributionForInstance(Instance exmp)
    throws Exception {

    // Extract the data
    Instances ins = exmp.relationalValue(1);
    if (m_Filter != null)
      ins = Filter.useFilter(ins, m_Filter);

    ins = Filter.useFilter(ins, m_Missing);

    int nI = ins.numInstances(), nA = ins.numAttributes();
    double[][] dat = new double [nI][nA];
    for (int j = 0; j < nI; j++){
      for (int k=0; k<nA; k++){
        dat[j][k] = ins.instance(j).value(k);
      }
    }
    //find the concept instance in the exemplar
    double min = Double.MAX_VALUE;
    double maxProb = -1.0;
    for (int j = 0; j < nI; j++){
      double exp = 0.0;
      for (int k = 0; k<nA; k++)  // for each attribute
        exp += (dat[j][k]-m_Par[k*2])*(dat[j][k]-m_Par[k*2])*m_Par[k*2+1]*m_Par[k*2+1];
      //the probability can be calculated as Math.exp(-exp)
      //to find the maximum Math.exp(-exp) is equivalent to find the minimum of (exp)
      if (exp < min)  {
        min     = exp;
        maxProb = Math.exp(-exp); //maximum probability of being positive   
      }
    }	

    // Compute the probability of the bag
    double[] distribution = new double[2];
    distribution[1] = maxProb; 
    distribution[0] = 1.0 - distribution[1];  //mininum prob. of being negative

    return distribution;
  }


  /**
   * Gets a string describing the classifier.
   *
   * @return a string describing the classifer built.
   */
  public String toString() {

    String result = "MIEMDD";
    if (m_Par == null) {
      return result + ": No model built yet.";
    }

    result += "\nCoefficients...\n"
      + "Variable       Point       Scale\n";
    for (int j = 0, idx=0; j < m_Par.length/2; j++, idx++) {
      result += m_Attributes.attribute(idx).name();
      result += " "+Utils.doubleToString(m_Par[j*2], 12, 4);
      result += " "+Utils.doubleToString(m_Par[j*2+1], 12, 4)+"\n";
    }

    return result;
  }

  /**
   * Main method for testing this class.
   *
   * @param argv should contain the command line arguments to the
   * scheme (see Evaluation)
   */
  public static void main(String[] argv) {
    runClassifier(new MIEMDD(), argv);
  }
}