logisticbase.java

一个数据挖掘软件ALPHAMINERR的整个过程的JAVA版源代码

	}
	
	// Evaluate / increment trainFs from the classifier
	for (int i = 0; i < trainFs.length; i++) {
	    double [] pred = new double [m_numClasses];
	    double predSum = 0;
	    for (int j = 0; j < m_numClasses; j++) {
		pred[j] = m_regressions[j][iteration]
		    .classifyInstance(trainNumeric.instance(i));
		predSum += pred[j];
	    }
	    predSum /= m_numClasses;
	    for (int j = 0; j < m_numClasses; j++) {
		trainFs[i][j] += (pred[j] - predSum) * (m_numClasses - 1) 
		    / m_numClasses;
	    }
	}
	
	// Compute the current probability estimates
	for (int i = 0; i < trainYs.length; i++) {
	    probs[i] = probs(trainFs[i]);
	}
	return true;
    }    

    /**
     * Helper function to initialize m_regressions.
     */
    protected SimpleLinearRegression[][] initRegressions(){
	SimpleLinearRegression[][] classifiers =   
	    new SimpleLinearRegression[m_numClasses][m_maxIterations];
	for (int j = 0; j < m_numClasses; j++) {
	    for (int i = 0; i < m_maxIterations; i++) {
		classifiers[j][i] = new SimpleLinearRegression();
		classifiers[j][i].setSuppressErrorMessage(true);
	    }
	}
	return classifiers;
    }

    /**
     * Converts training data to numeric version. The class variable is replaced by a pseudo-class 
     * used by LogitBoost.
     */
    protected Instances getNumericData(Instances data) throws Exception{
	Instances numericData = new Instances(data);
	
	int classIndex = numericData.classIndex();
	numericData.setClassIndex(-1);
	numericData.deleteAttributeAt(classIndex);
	numericData.insertAttributeAt(new Attribute("'pseudo class'"), classIndex);
	numericData.setClassIndex(classIndex);
	return numericData;
    }
    
    /**
     * Helper function for cutting back m_regressions to the set of classifiers (corresponsing to the number of 
     * LogitBoost iterations) that gave the smallest error.
     */
    protected SimpleLinearRegression[][] selectRegressions(SimpleLinearRegression[][] classifiers){
	SimpleLinearRegression[][] goodClassifiers = 
	    new SimpleLinearRegression[m_numClasses][m_numRegressions];
	
	for (int j = 0; j < m_numClasses; j++) {
	    for (int i = 0; i < m_numRegressions; i++) {
		goodClassifiers[j][i] = classifiers[j][i];
	    }
	}
	return goodClassifiers;
    }		
    
    /**
     * Computes the LogitBoost response variable from y/p values (actual/estimated class probabilities).
     */
    protected double getZ(double actual, double p) {
	double z;
	if (actual == 1) {
	    z = 1.0 / p;
	    if (z > Z_MAX) { // threshold
		z = Z_MAX;
	    }
	} else {
	    z = -1.0 / (1.0 - p);
	    if (z < -Z_MAX) { // threshold
		z = -Z_MAX;
	    }
	}
	return z;
    }
    
    /**
     * Computes the LogitBoost response for an array of y/p values (actual/estimated class probabilities).
     */
    protected double[][] getZs(double[][] probs, double[][] dataYs) {
	
	double[][] dataZs = new double[probs.length][m_numClasses];
	for (int j = 0; j < m_numClasses; j++) 
	    for (int i = 0; i < probs.length; i++) dataZs[i][j] = getZ(dataYs[i][j], probs[i][j]);
	return dataZs;
    }
    
    /**
     * Computes the LogitBoost weights from an array of y/p values (actual/estimated class probabilities).
     */
    protected double[][] getWs(double[][] probs, double[][] dataYs) {
	
	double[][] dataWs = new double[probs.length][m_numClasses];
	for (int j = 0; j < m_numClasses; j++) 
	    for (int i = 0; i < probs.length; i++){
	    double z = getZ(dataYs[i][j], probs[i][j]);
	    dataWs[i][j] = (dataYs[i][j] - probs[i][j]) / z;
	    }
	return dataWs;
    }

    /**
     * Computes the p-values (probabilities for the classes) from the F-values of the logistic model.
     */
    protected double[] probs(double[] Fs) {
	
	double maxF = -Double.MAX_VALUE;
	for (int i = 0; i < Fs.length; i++) {
	    if (Fs[i] > maxF) {
		maxF = Fs[i];
	    }
	}   
	double sum = 0;
	double[] probs = new double[Fs.length];
	for (int i = 0; i < Fs.length; i++) {
	    probs[i] = Math.exp(Fs[i] - maxF);   
	    sum += probs[i];
	}
	
	Utils.normalize(probs, sum);
	return probs;
    }

    /**
     * Computes the Y-values (actual class probabilities) for a set of instances.
     */
    protected double[][] getYs(Instances data){
	
	double [][] dataYs = new double [data.numInstances()][m_numClasses];
	for (int j = 0; j < m_numClasses; j++) {
	    for (int k = 0; k < data.numInstances(); k++) {
		dataYs[k][j] = (data.instance(k).classValue() == j) ? 
		    1.0: 0.0;
	    }
	}
	return dataYs;
    }

    /**
     * Computes the F-values for a single instance.
     */
    protected double[] getFs(Instance instance) throws Exception{
	
	double [] pred = new double [m_numClasses];
	double [] instanceFs = new double [m_numClasses]; 
	
	//add up the predictions from the simple regression functions
	for (int i = 0; i < m_numRegressions; i++) {
	    double predSum = 0;
	    for (int j = 0; j < m_numClasses; j++) {
		pred[j] = m_regressions[j][i].classifyInstance(instance);
		predSum += pred[j];
	    }
	    predSum /= m_numClasses;
	    for (int j = 0; j < m_numClasses; j++) {
		instanceFs[j] += (pred[j] - predSum) * (m_numClasses - 1) 
		    / m_numClasses;
	    }
	}	
	
	return instanceFs; 
    } 
    
    /**
     * Computes the F-values for a set of instances.
     */
    protected double[][] getFs(Instances data) throws Exception{
	
	double[][] dataFs = new double[data.numInstances()][];
       
	for (int k = 0; k < data.numInstances(); k++) {
	    dataFs[k] = getFs(data.instance(k));
	}
	
	return dataFs;	
    }   

    /**
     * Computes the p-values (probabilities for the different classes) from the F-values for a set of instances.
     */
    protected double[][] getProbs(double[][] dataFs){
	
	int numInstances = dataFs.length;
	double[][] probs = new double[numInstances][];
	
	for (int k = 0; k < numInstances; k++) {       
	    probs[k] = probs(dataFs[k]);
	}
	return probs;
    }
    
    /**
     * Returns the likelihood of the Y-values (actual class probabilities) given the 
     * p-values (current probability estimates).
     */
    protected double logLikelihood(double[][] dataYs, double[][] probs) {
	
	double logLikelihood = 0;
	for (int i = 0; i < dataYs.length; i++) {
	    for (int j = 0; j < m_numClasses; j++) {
		if (dataYs[i][j] == 1.0) {
		    logLikelihood -= Math.log(probs[i][j]);
		}
	    }
	}
	return logLikelihood / (double)dataYs.length;
    }

    /**
     * Returns an array of the indices of the attributes used in the logistic model.
     * The first dimension is the class, the second dimension holds a list of attribute indices.
     * Attribute indices start at zero.
     * @return the array of attribute indices
     */
    public int[][] getUsedAttributes(){
	
	int[][] usedAttributes = new int[m_numClasses][];
	
	//first extract coefficients
	double[][] coefficients = getCoefficients();
	
	for (int j = 0; j < m_numClasses; j++){
	    
	    //boolean array indicating if attribute used
	    boolean[] attributes = new boolean[m_numericDataHeader.numAttributes()];
	    for (int i = 0; i < attributes.length; i++) {
		//attribute used if coefficient > 0
		if (!Utils.eq(coefficients[j][i + 1],0)) attributes[i] = true;
	    }
	    	    
	    int numAttributes = 0;
	    for (int i = 0; i < m_numericDataHeader.numAttributes(); i++) if (attributes[i]) numAttributes++;
	    
	    //"collect" all attributes into array of indices
	    int[] usedAttributesClass = new int[numAttributes];
	    int count = 0;
	    for (int i = 0; i < m_numericDataHeader.numAttributes(); i++) {
		if (attributes[i]) {
		usedAttributesClass[count] = i;
		count++;
		} 
	    }
	    
	    usedAttributes[j] = usedAttributesClass;
	}
	
	return usedAttributes;
    }

    /**
     * The number of LogitBoost iterations performed (= the number of simple regression functions fit).
     */
    public int getNumRegressions() {
	return m_numRegressions;
    }

    /**
     * Sets the parameter "maxIterations".
     */
    public void setMaxIterations(int maxIterations) {
	m_maxIterations = maxIterations;
    }
    
    /**
     * Sets the option "heuristicStop".
     */
    public void setHeuristicStop(int heuristicStop){
	m_heuristicStop = heuristicStop;
    }

    /**
     * Returns the maxIterations parameter.
     */
    public int getMaxIterations(){
	return m_maxIterations;
    }
        
    /**
     * Returns an array holding the coefficients of the logistic model.
     * First dimension is the class, the second one holds a list of coefficients.
     * At position zero, the constant term of the model is stored, then, the coefficients for
     * the attributes in ascending order.
     * @return the array of coefficients
     */
    protected double[][] getCoefficients(){
	double[][] coefficients = new double[m_numClasses][m_numericDataHeader.numAttributes() + 1];
	for (int j = 0; j < m_numClasses; j++) {
	    //go through simple regression functions and add their coefficient to the coefficient of
	    //the attribute they are built on.
	    for (int i = 0; i < m_numRegressions; i++) {
		
		double slope = m_regressions[j][i].getSlope();
		double intercept = m_regressions[j][i].getIntercept();
		int attribute = m_regressions[j][i].getAttributeIndex();
		
		coefficients[j][0] += intercept;
		coefficients[j][attribute + 1] += slope;
	    }
	}
	return coefficients;
    }

    /**
     * Returns the fraction of all attributes in the data that are used in the logistic model (in percent).
     * An attribute is used in the model if it is used in any of the models for the different classes.
     */
    public double percentAttributesUsed(){	
	boolean[] attributes = new boolean[m_numericDataHeader.numAttributes()];
	
	double[][] coefficients = getCoefficients();
	for (int j = 0; j < m_numClasses; j++){
	    for (int i = 1; i < m_numericDataHeader.numAttributes() + 1; i++) {
		//attribute used if it is used in any class, note coefficients are shifted by one (because
		//of constant term).
		if (!Utils.eq(coefficients[j][i],0)) attributes[i - 1] = true;
	    }
	}
	
	//count number of used attributes (without the class attribute)
	double count = 0;
	for (int i = 0; i < attributes.length; i++) if (attributes[i]) count++;
	return count / (double)(m_numericDataHeader.numAttributes() - 1) * 100.0;
    }
    
    /**
     * Returns a description of the logistic model (i.e., attributes and coefficients).
     */
    public String toString(){
	
	StringBuffer s = new StringBuffer();	

	//get used attributes
	int[][] attributes = getUsedAttributes();
	
	//get coefficients
	double[][] coefficients = getCoefficients();
	
	for (int j = 0; j < m_numClasses; j++) {
	    s.append("\nClass "+j+" :\n");
	    //constant term
	    s.append(Utils.doubleToString(coefficients[j][0],4,2)+" + \n");
	    for (int i = 0; i < attributes[j].length; i++) {		
		//attribute/coefficient pairs
		s.append("["+m_numericDataHeader.attribute(attributes[j][i]).name()+"]");
		s.append(" * " + Utils.doubleToString(coefficients[j][attributes[j][i]+1],4,2));
		if (i != attributes[j].length - 1) s.append(" +");
		s.append("\n");	    
	    }
	}	
	return new String(s);
    }

    /** 
     * Returns class probabilities for an instance.
     *
     * @exception Exception if distribution can't be computed successfully
     */
    public double[] distributionForInstance(Instance instance) throws Exception {
	
	instance = (Instance)instance.copy();	

	//set to numeric pseudo-class
      	instance.setDataset(m_numericDataHeader);		
	
	//calculate probs via Fs
	return probs(getFs(instance));
    }

    /**
     * Cleanup in order to save memory.
     */
    public void cleanup() {
	//save just header info
	m_train = new Instances(m_train,0);
	m_numericData = null;	
    }
}