/*
* RapidMiner
*
* Copyright (C) 2001-2011 by Rapid-I and the contributors
*
* Complete list of developers available at our web site:
*
* http://rapid-i.com
*
* This program is free software: you can redistribute it and/or modify
* it under the terms of the GNU Affero General Public License as published by
* the Free Software Foundation, either version 3 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 Affero General Public License for more details.
*
* You should have received a copy of the GNU Affero General Public License
* along with this program. If not, see http://www.gnu.org/licenses/.
*/
package com.rapidminer.operator.learner.bayes;
import java.util.ArrayList;
import java.util.Collection;
import com.rapidminer.example.Attribute;
import com.rapidminer.example.Example;
import com.rapidminer.example.ExampleSet;
import com.rapidminer.tools.Tools;
import com.rapidminer.tools.math.distribution.DiscreteDistribution;
import com.rapidminer.tools.math.distribution.Distribution;
import com.rapidminer.tools.math.distribution.kernel.FullKernelDistribution;
import com.rapidminer.tools.math.distribution.kernel.GreedyKernelDistribution;
import com.rapidminer.tools.math.distribution.kernel.KernelDistribution;
/**
* KernelDistributionModel is a model for learners which estimate
* distributions of attribute values from example sets like NaiveBayes.
*
* Predictions are calculated as product of the conditional probabilities
* for all attributes times the class probability.
*
* The basic learning concept is to simply count occurrences of classes
* and attribute values. This means no probabilities are calculated during
* the learning step. This is only done before output. Optionally, this
* calculation can apply a Laplace correction which means in particular
* that zero probabilities are avoided which would hide information
* in distributions of other attributes.
*
* @author Tobias Malbrecht
*/
public class KernelDistributionModel extends DistributionModel {
private static final long serialVersionUID = -402827845291958569L;
private static final String UNKNOWN_VALUE_NAME = "unknown";
/** The number of classes. */
private int numberOfClasses;
/** The number of attributes. */
private int numberOfAttributes;
/** Flags indicating which attribute is nominal. */
private boolean[] nominal;
/** Class name (used for result displaying). */
private String className;
/** Class values (used for result displaying). */
private String[] classValues;
/** Attribute names (used for result displaying). */
private String[] attributeNames;
/** Nominal attribute values (used for result displaying). */
private String[][] attributeValues;
/** Total weight (or number) of examples used to build the model. */
private double totalWeight;
/** Total weight of examples belonging to the separate classes. */
private double[] classWeights;
/**
* Specifies the total weight of examples in which the different combinations
* of classes and (nominal) attribute values co-occur. In the case of numeric
* attributes the (weighted) sum and the (weighted) sum of the squared
* attribute values are stored which are needed to calculate the mean and the
* standard deviation/variance of the resulting (assumed) normal distribution.
*
* Array dimensions:
* 1st: attributes
* 2nd: classes
* 3nd: nominal values or value sum (index=0) and squared value sum (index=1)
*/
private double[][][] weightSums;
/** Class log (!) a-priori probabilities. */
private double[] priors;
/**
* Specifies the a-postiori distributions. Contains the log (!) a-postiori
* Probabilities that certain values occur given the class value for nominal
* values. Contains the means and standard deviations for numerical attributes.
*
* Array dimensions:
* 1st: attributes
* 2nd: classes
* 3nd: nominal values or mean (index=0) and standard deviation (index=1)
*/
private double[][][] distributionProperties;
/**
* The kernel distributions for the nominal attributes.
*/
private KernelDistribution[][] kernelDistributions;
/**
* Captures if laplace correction should be applied when calculating
* probabilities.
*/
boolean laplaceCorrectionEnabled;
/**
* Indicates if the model has recently been updated and the actual probabilities
* have to be calculated.
*/
private boolean modelRecentlyUpdated;
private boolean useApplianceGrid;
private double[][][] grid;
private int gridSize = 200;
public KernelDistributionModel(ExampleSet exampleSet, boolean laplaceCorrectionEnabled, int estimationMode, int bandwidthSelectionMode, double bandwidth, int numberOfKernels, int gridSize) {
super(exampleSet);
this.laplaceCorrectionEnabled = laplaceCorrectionEnabled;
this.useApplianceGrid = (gridSize > 10);
this.gridSize = gridSize;
Attribute labelAttribute = exampleSet.getAttributes().getLabel();
numberOfClasses = labelAttribute.getMapping().size();
numberOfAttributes = exampleSet.getAttributes().size();
nominal = new boolean[numberOfAttributes];
attributeNames = new String[numberOfAttributes];
attributeValues = new String[numberOfAttributes][];
className = labelAttribute.getName();
classValues = new String[numberOfClasses];
for (int i = 0; i < numberOfClasses; i++) {
classValues[i] = labelAttribute.getMapping().mapIndex(i);
}
int attributeIndex = 0;
weightSums = new double[numberOfAttributes][numberOfClasses][];
distributionProperties = new double[numberOfAttributes][numberOfClasses][];
kernelDistributions = new KernelDistribution[numberOfAttributes][numberOfClasses];
for (Attribute attribute : exampleSet.getAttributes()) {
attributeNames[attributeIndex] = attribute.getName();
if (attribute.isNominal()) {
nominal[attributeIndex] = true;
int mappingSize = attribute.getMapping().size() + 1;
attributeValues[attributeIndex] = new String[mappingSize];
for (int i = 0; i < mappingSize - 1; i++) {
attributeValues[attributeIndex][i] = attribute.getMapping().mapIndex(i);
}
attributeValues[attributeIndex][mappingSize - 1] = UNKNOWN_VALUE_NAME;
for (int i = 0; i < numberOfClasses; i++) {
weightSums[attributeIndex][i] = new double[mappingSize];
distributionProperties[attributeIndex][i] = new double[mappingSize];
}
} else {
nominal[attributeIndex] = false;
for (int i = 0; i < numberOfClasses; i++) {
switch (estimationMode) {
case KernelNaiveBayes.ESTIMATION_MODE_FULL:
switch (bandwidthSelectionMode) {
case KernelNaiveBayes.BANDWIDTH_SELECTION_MODE_HEURISTIC:
kernelDistributions[attributeIndex][i] = new FullKernelDistribution();
break;
case KernelNaiveBayes.BANDWIDTH_SELECTION_MODE_FIX:
kernelDistributions[attributeIndex][i] = new FullKernelDistribution(bandwidth);
break;
}
break;
case KernelNaiveBayes.ESTIMATION_MODE_GREEDY:
kernelDistributions[attributeIndex][i] = new GreedyKernelDistribution(bandwidth, numberOfKernels);
break;
default:
kernelDistributions[attributeIndex][i] = new FullKernelDistribution();
}
}
}
attributeIndex++;
}
// initialization of total and a priori weight counters
totalWeight = 0.0d;
classWeights = new double[numberOfClasses];
priors = new double[numberOfClasses];
if (useApplianceGrid) {
grid = new double[numberOfAttributes][numberOfClasses][];
}
// update the model
update(exampleSet);
// calculate the probabilities
updateDistributionProperties();
}
@Override
public String[] getAttributeNames() {
return this.attributeNames;
}
@Override
public int getNumberOfAttributes() {
return this.attributeNames.length;
}
/**
* Updates the model by counting the occurrences of classes and attribute values
* in combination with the class values.
*
* ATTENTION: only updates the weight counters, distribution properties are not
* updated, call updateDistributionProperties() to accomplish this task
*/
@Override
public void update(ExampleSet exampleSet) {
Attribute weightAttribute = exampleSet.getAttributes().getWeight();
for (Example example : exampleSet) {
double weight = weightAttribute == null ? 1.0d : example.getWeight();
totalWeight += weight;
double labelValue = example.getLabel();
if (!Double.isNaN(labelValue)) {
int classIndex = (int) example.getLabel();
classWeights[classIndex] += weight;
int attributeIndex = 0;
for (Attribute attribute : exampleSet.getAttributes()) {
double attributeValue = example.getValue(attribute);
if (nominal[attributeIndex]) {
if (!Double.isNaN(attributeValue)) {
if ((int) attributeValue < weightSums[attributeIndex][classIndex].length - 1) {
weightSums[attributeIndex][classIndex][(int) attributeValue] += weight;
} else {
// extend weight array if attribute value is not in mapping
for (int i = 0; i < numberOfClasses; i++) {
double[] newWeightSums = new double[(int) attributeValue + 2];
newWeightSums[newWeightSums.length - 1] = weightSums[attributeIndex][i][weightSums[attributeIndex][i].length - 1];
for (int j = 0; j < weightSums[attributeIndex][i].length - 1; j++) {
newWeightSums[j] = weightSums[attributeIndex][i][j];
}
weightSums[attributeIndex][i] = newWeightSums;
distributionProperties[attributeIndex][i] = new double[(int) attributeValue + 2];
}
weightSums[attributeIndex][classIndex][(int) attributeValue] += weight;
// recreate internal attribute value mapping
attributeValues[attributeIndex] = new String[(int) attributeValue + 2];
for (int i = 0; i < attributeValues[attributeIndex].length - 1; i++) {
attributeValues[attributeIndex][i] = attribute.getMapping().mapIndex(i);
}
attributeValues[attributeIndex][attributeValues[attributeIndex].length - 1] = UNKNOWN_VALUE_NAME;
}
} else {
weightSums[attributeIndex][classIndex][weightSums[attributeIndex][classIndex].length - 1] += weight;
}
} else {
if (!Double.isNaN(attributeValue)) {
kernelDistributions[attributeIndex][classIndex].update(attributeValue, weight);
}
}
attributeIndex++;
}
}
}
modelRecentlyUpdated = true;
}
/**
* Updates the distribution properties by calculating the logged probabilities
* and distribution parameters on the basis of the weight counters.
*/
private void updateDistributionProperties() {
double f = laplaceCorrectionEnabled ? 1 / totalWeight : Double.MIN_VALUE;
for (int i = 0; i < numberOfClasses; i++) {
priors[i] = Math.log(classWeights[i] / totalWeight);
}
for (int i = 0; i < numberOfAttributes; i++) {
if (nominal[i]) {
for (int j = 0; j < numberOfClasses; j++) {
for (int k = 0; k < weightSums[i][j].length; k++) {
distributionProperties[i][j][k] = Math.log((weightSums[i][j][k] + f) / (classWeights[j] + f * weightSums[i][j].length));
}
}
}
}
if (useApplianceGrid) {
for (int i = 0; i < numberOfClasses; i++) {
for (int j = 0; j < numberOfAttributes; j++) {
if (!nominal[j]) {
double lowerBound = kernelDistributions[j][i].getLowerBound();
double upperBound = kernelDistributions[j][i].getUpperBound();
double precision = (upperBound - lowerBound) / gridSize;
grid[j][i] = new double[gridSize + 1];
for (int k = 0; k < gridSize + 1; k++) {
grid[j][i][k] = kernelDistributions[j][i].getProbability(lowerBound + k * precision);
}
}
}
}
}
modelRecentlyUpdated = false;
}
/**
* Perform predictions based on the distribution properties.
*/
@Override
public ExampleSet performPrediction(ExampleSet exampleSet, Attribute predictedLabel) {
if (modelRecentlyUpdated) {
updateDistributionProperties();
}
for (Example example : exampleSet) {
double[] probabilities = new double[numberOfClasses];
double maxLogProbability = Double.NEGATIVE_INFINITY;
int mostProbableClass = 0;
double probabilitySum = 0;
for (int i = 0; i < numberOfClasses; i++) {
double logProbability = priors[i];
int j = 0;
for (Attribute attribute : exampleSet.getAttributes()) {
double value = example.getValue(attribute);
if (nominal[j]) {
if (!Double.isNaN(value) && (int) value < distributionProperties[j][i].length) {
logProbability += distributionProperties[j][i][(int) value];
} else {
logProbability += distributionProperties[j][i][distributionProperties[j][i].length - 1];
}
} else {
if (!Double.isNaN(value)) {
if (useApplianceGrid) {
double upperBound = kernelDistributions[j][i].getUpperBound();
double lowerBound = kernelDistributions[j][i].getLowerBound();
double precision = (upperBound - lowerBound) / gridSize;
if (value >= lowerBound && value <= kernelDistributions[j][i].getUpperBound()) {
logProbability += Math.log(grid[j][i][(int) ((value - lowerBound) / precision)]);
} else {
logProbability += Math.log(kernelDistributions[j][i].getProbability(value));
}
} else {
logProbability += Math.log(kernelDistributions[j][i].getProbability(value));
}
}
}
j++;
}
if (!Double.isNaN(logProbability) && logProbability > maxLogProbability) {
maxLogProbability = logProbability;
mostProbableClass = i;
}
probabilities[i] = logProbability;
}
for (int i = 0; i < numberOfClasses; i++) {
if (!Double.isNaN(probabilities[i])) {
probabilities[i] = Math.exp(probabilities[i] - maxLogProbability);
probabilitySum += probabilities[i];
} else {
probabilities[i] = 0;
}
}
if (maxLogProbability == Double.NEGATIVE_INFINITY) {
example.setPredictedLabel(Double.NaN);
for (int i = 0; i < numberOfClasses; i++) {
example.setConfidence(classValues[i], Double.NaN);
}
} else {
example.setPredictedLabel(mostProbableClass);
for (int i = 0; i < numberOfClasses; i++) {
example.setConfidence(classValues[i], probabilities[i] / probabilitySum);
}
}
}
return exampleSet;
}
public void setLaplaceCorrectionEnabled(boolean laplaceCorrectionEnabled) {
this.laplaceCorrectionEnabled = laplaceCorrectionEnabled;
}
public boolean getLaplaceCorrectionEnabled() {
return laplaceCorrectionEnabled;
}
@Override
public double getLowerBound(int attributeIndex) {
if (!nominal[attributeIndex]) {
double lowerBound = Double.POSITIVE_INFINITY;
for (int i = 0; i < numberOfClasses; i++) {
double currentLowerBound = kernelDistributions[attributeIndex][i].getLowerBound();
if (!Double.isNaN(currentLowerBound)) {
lowerBound = Math.min(lowerBound, currentLowerBound);
}
}
return lowerBound;
} else {
return Double.NaN;
}
}
@Override
public double getUpperBound(int attributeIndex) {
if (!nominal[attributeIndex]) {
double upperBound = Double.NEGATIVE_INFINITY;
for (int i = 0; i < numberOfClasses; i++) {
double currentUpperBound = kernelDistributions[attributeIndex][i].getUpperBound();
if (!Double.isNaN(currentUpperBound)) {
upperBound = Math.max(upperBound, currentUpperBound);
}
}
return upperBound;
} else {
return Double.NaN;
}
}
@Override
public boolean isDiscrete(int attributeIndex) {
if (attributeIndex>= 0 && attributeIndex < nominal.length) {
return nominal[attributeIndex];
}
return false;
}
@Override
public Collection<Integer> getClassIndices() {
Collection<Integer> classValueIndices = new ArrayList<Integer>(numberOfClasses);
for (int i = 0; i < numberOfClasses; i++) {
classValueIndices.add(i);
}
return classValueIndices;
}
@Override
public int getNumberOfClasses() {
return numberOfClasses;
}
@Override
public String getClassName(int index) {
return classValues[index];
}
@Override
public Distribution getDistribution(int classIndex, int attributeIndex) {
if (nominal[attributeIndex]) {
double[] probabilities = new double[distributionProperties[attributeIndex][classIndex].length];
for (int i = 0; i < probabilities.length; i++) {
probabilities[i] = Math.exp(distributionProperties[attributeIndex][classIndex][i]);
}
return new DiscreteDistribution(attributeNames[attributeIndex], probabilities, attributeValues[attributeIndex]);
} else {
return kernelDistributions[attributeIndex][classIndex];
}
}
@Override
public String toString() {
if (modelRecentlyUpdated) {
updateDistributionProperties();
}
StringBuffer buffer = new StringBuffer();
buffer.append("Distribution model for label attribute " + className);
buffer.append(Tools.getLineSeparators(2));
for (int i = 0; i < numberOfClasses; i++) {
String classTitle = "Class " + classValues[i] + " (" + Tools.formatNumber(Math.exp(priors[i])) + ")";
buffer.append(Tools.getLineSeparator());
buffer.append(classTitle);
buffer.append(Tools.getLineSeparator());
buffer.append(attributeNames.length + " distributions");
buffer.append(Tools.getLineSeparator());
}
return buffer.toString();
}
}