/*
* 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.
*/
/*
* NNge.java
* Copyright (C) 2002 Brent Martin
*
*/
package weka.classifiers.rules;
import weka.classifiers.Classifier;
import weka.classifiers.UpdateableClassifier;
import weka.core.Capabilities;
import weka.core.Instance;
import weka.core.Instances;
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.LinkedList;
import java.util.Vector;
import weka.classifiers.AbstractClassifier;
import weka.core.DenseInstance;
/**
<!-- globalinfo-start -->
* Nearest-neighbor-like algorithm using non-nested generalized exemplars (which are hyperrectangles that can be viewed as if-then rules). For more information, see <br/>
* <br/>
* Brent Martin (1995). Instance-Based learning: Nearest Neighbor With Generalization. Hamilton, New Zealand.<br/>
* <br/>
* Sylvain Roy (2002). Nearest Neighbor With Generalization. Christchurch, New Zealand.
* <p/>
<!-- globalinfo-end -->
*
<!-- technical-bibtex-start -->
* BibTeX:
* <pre>
* @mastersthesis{Martin1995,
* address = {Hamilton, New Zealand},
* author = {Brent Martin},
* school = {University of Waikato},
* title = {Instance-Based learning: Nearest Neighbor With Generalization},
* year = {1995}
* }
*
* @unpublished{Roy2002,
* address = {Christchurch, New Zealand},
* author = {Sylvain Roy},
* school = {University of Canterbury},
* title = {Nearest Neighbor With Generalization},
* year = {2002}
* }
* </pre>
* <p/>
<!-- technical-bibtex-end -->
*
<!-- options-start -->
* Valid options are: <p/>
*
* <pre> -G <value>
* Number of attempts of generalisation.
* </pre>
*
* <pre> -I <value>
* Number of folder for computing the mutual information.
* </pre>
*
<!-- options-end -->
*
* @author Brent Martin (bim20@cosc.canterbury.ac.nz)
* @author Sylvain Roy (sro33@student.canterbury.ac.nz)
* @version $Revision: 5529 $
*/
public class NNge
extends AbstractClassifier
implements UpdateableClassifier, OptionHandler, TechnicalInformationHandler {
/** for serialization */
static final long serialVersionUID = 4084742275553788972L;
/**
* Returns a string describing classifier
* @return a description suitable for
* displaying in the explorer/experimenter gui
*/
public String globalInfo() {
return "Nearest-neighbor-like algorithm using non-nested generalized exemplars "
+ "(which are hyperrectangles that can be viewed as if-then rules). 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;
TechnicalInformation additional;
result = new TechnicalInformation(Type.MASTERSTHESIS);
result.setValue(Field.AUTHOR, "Brent Martin");
result.setValue(Field.YEAR, "1995");
result.setValue(Field.TITLE, "Instance-Based learning: Nearest Neighbor With Generalization");
result.setValue(Field.SCHOOL, "University of Waikato");
result.setValue(Field.ADDRESS, "Hamilton, New Zealand");
additional = result.add(Type.UNPUBLISHED);
additional.setValue(Field.AUTHOR, "Sylvain Roy");
additional.setValue(Field.YEAR, "2002");
additional.setValue(Field.TITLE, "Nearest Neighbor With Generalization");
additional.setValue(Field.SCHOOL, "University of Canterbury");
additional.setValue(Field.ADDRESS, "Christchurch, New Zealand");
return result;
}
/**
* Implements Exemplar as used by NNge : parallel axis hyperrectangle.
*/
private class Exemplar
extends Instances {
/** for serialization */
static final long serialVersionUID = 3960180128928697216L;
/** List of all the Exemplar */
private Exemplar previous = null;
private Exemplar next = null;
/** List of all the Exemplar with the same class */
private Exemplar previousWithClass = null;
private Exemplar nextWithClass = null;
/** The NNge which owns this Exemplar */
private NNge m_NNge;
/** class of the Exemplar */
private double m_ClassValue;
/** Number of correct prediction for this examplar */
private int m_PositiveCount = 1;
/** Number of incorrect prediction for this examplar */
private int m_NegativeCount = 0;
/** The max borders of the rectangle for numeric attributes */
private double[] m_MaxBorder;
/** The min borders of the rectangle for numeric attributes */
private double[] m_MinBorder;
/** The ranges of the hyperrectangle for nominal attributes */
private boolean[][] m_Range;
/** the arrays used by preGeneralise */
private double[] m_PreMaxBorder = null;
private double[] m_PreMinBorder = null;
private boolean[][] m_PreRange = null;
private Instance m_PreInst = null;
/**
* Build a new empty Exemplar
*
* @param nnge the classifier which owns this Exemplar
* @param inst the instances from which the header information is to be taken
* @param size the capacity of the Exemplar
* @param classV the class of the Exemplar
*/
private Exemplar (NNge nnge, Instances inst, int size, double classV){
super(inst, size);
m_NNge = nnge;
m_ClassValue = classV;
m_MinBorder = new double[numAttributes()];
m_MaxBorder = new double[numAttributes()];
m_Range = new boolean[numAttributes()][];
for(int i = 0; i < numAttributes(); i++){
if(attribute(i).isNumeric()){
m_MinBorder[i] = Double.POSITIVE_INFINITY;
m_MaxBorder[i] = Double.NEGATIVE_INFINITY;
m_Range[i] = null;
} else {
m_MinBorder[i] = Double.NaN;
m_MaxBorder[i] = Double.NaN;
m_Range[i] = new boolean[attribute(i).numValues() + 1];
for(int j = 0; j < attribute(i).numValues() + 1; j++){
m_Range[i][j] = false;
}
}
}
}
/**
* Generalise the Exemplar with inst
*
* @param inst the new example used for the generalisation
* @throws Exception if either the class of inst is not equal to the class of the Exemplar or inst misses a value.
*/
private void generalise(Instance inst) throws Exception {
if(m_ClassValue != inst.classValue())
throw new Exception("Exemplar.generalise : Incompatible instance's class.");
add(inst);
/* extends each range in order to cover inst */
for(int i = 0; i < numAttributes(); i++){
if(inst.isMissing(i))
throw new Exception("Exemplar.generalise : Generalisation with missing feature impossible.");
if(i == classIndex())
continue;
if(attribute(i).isNumeric()){
if(m_MaxBorder[i] < inst.value(i))
m_MaxBorder[i] = inst.value(i);
if(inst.value(i) < m_MinBorder[i])
m_MinBorder[i] = inst.value(i);
} else {
m_Range[i][(int) inst.value(i)] = true;
}
}
}
/**
* pre-generalise the Exemplar with inst
* i.e. the boundaries of the Exemplar include inst but the Exemplar still doesn't 'own' inst.
* To be complete, the generalisation must be validated with validateGeneralisation.
* the generalisation can be canceled with cancelGeneralisation.
* @param inst the new example used for the generalisation
* @throws Exception if either the class of inst is not equal to the class of the Exemplar or inst misses a value.
*/
private void preGeneralise(Instance inst) throws Exception {
if(m_ClassValue != inst.classValue())
throw new Exception("Exemplar.preGeneralise : Incompatible instance's class.");
m_PreInst = inst;
/* save the current state */
m_PreRange = new boolean[numAttributes()][];
m_PreMinBorder = new double[numAttributes()];
m_PreMaxBorder = new double[numAttributes()];
for(int i = 0; i < numAttributes(); i++){
if(attribute(i).isNumeric()){
m_PreMinBorder[i] = m_MinBorder[i];
m_PreMaxBorder[i] = m_MaxBorder[i];
} else {
m_PreRange[i] = new boolean[attribute(i).numValues() + 1];
for(int j = 0; j < attribute(i).numValues() + 1; j++){
m_PreRange[i][j] = m_Range[i][j];
}
}
}
/* perform the pre-generalisation */
for(int i = 0; i < numAttributes(); i++){
if(inst.isMissing(i))
throw new Exception("Exemplar.preGeneralise : Generalisation with missing feature impossible.");
if(i == classIndex())
continue;
if(attribute(i).isNumeric()){
if(m_MaxBorder[i] < inst.value(i))
m_MaxBorder[i] = inst.value(i);
if(inst.value(i) < m_MinBorder[i])
m_MinBorder[i] = inst.value(i);
} else {
m_Range[i][(int) inst.value(i)] = true;
}
}
}
/**
* Validates a generalisation started with preGeneralise.
* Watch out, preGeneralise must have been called before.
*
* @throws Exception is thrown if preGeneralise hasn't been called before
*/
private void validateGeneralisation() throws Exception {
if(m_PreInst == null){
throw new Exception("Exemplar.validateGeneralisation : validateGeneralisation called without previous call to preGeneralise!");
}
add(m_PreInst);
m_PreRange = null;
m_PreMinBorder = null;
m_PreMaxBorder = null;
}
/**
* Cancels a generalisation started with preGeneralise.
* Watch out, preGeneralise must have been called before.
*
* @throws Exception is thrown if preGeneralise hasn't been called before
*/
private void cancelGeneralisation() throws Exception {
if(m_PreInst == null){
throw new Exception("Exemplar.cancelGeneralisation : cancelGeneralisation called without previous call to preGeneralise!");
}
m_PreInst = null;
m_Range = m_PreRange;
m_MinBorder = m_PreMinBorder;
m_MaxBorder = m_PreMaxBorder;
m_PreRange = null;
m_PreMinBorder = null;
m_PreMaxBorder = null;
}
/**
* return true if inst is held by this Exemplar, false otherwise
*
* @param inst an Instance
* @return true if inst is held by this hyperrectangle, false otherwise
*/
private boolean holds(Instance inst) {
if(numInstances() == 0)
return false;
for(int i = 0; i < numAttributes(); i++){
if(i != classIndex() && !holds(i, inst.value(i)))
return false;
}
return true;
}
/**
* return true if value is inside the Exemplar along the attrIndex attribute.
*
* @param attrIndex the index of an attribute
* @param value a value along the attrIndexth attribute
* @return true if value is inside the Exemplar along the attrIndex attribute.
*/
private boolean holds(int attrIndex, double value) {
if (numAttributes() == 0)
return false;
if(attribute(attrIndex).isNumeric())
return(m_MinBorder[attrIndex] <= value && value <= m_MaxBorder[attrIndex]);
else
return m_Range[attrIndex][(int) value];
}
/**
* Check if the Examplar overlaps ex
*
* @param ex an Exemplar
* @return true if ex is overlapped by the Exemplar
* @throws Exception
*/
private boolean overlaps(Exemplar ex) {
if(ex.isEmpty() || isEmpty())
return false;
for (int i = 0; i < numAttributes(); i++){
if(i == classIndex()){
continue;
}
if (attribute(i).isNumeric() &&
(ex.m_MaxBorder[i] < m_MinBorder[i] || ex.m_MinBorder[i] > m_MaxBorder[i])){
return false;
}
if (attribute(i).isNominal()) {
boolean in = false;
for (int j = 0; j < attribute(i).numValues() + 1; j++){
if(m_Range[i][j] && ex.m_Range[i][j]){
in = true;
break;
}
}
if(!in) return false;
}
}
return true;
}
/**
* Compute the distance between the projection of inst and this Exemplar along the attribute attrIndex.
* If inst misses its value along the attribute, the function returns 0.
*
* @param inst an instance
* @param attrIndex the index of the attribute
* @return the distance between the projection of inst and this Exemplar along the attribute attrIndex.
*/
private double attrDistance(Instance inst, int attrIndex) {
if(inst.isMissing(attrIndex))
return 0;
/* numeric attribute */
if(attribute(attrIndex).isNumeric()){
double norm = m_NNge.m_MaxArray[attrIndex] - m_NNge.m_MinArray[attrIndex];
if(norm <= 0)
norm = 1;
if (m_MaxBorder[attrIndex] < inst.value(attrIndex)) {
return (inst.value(attrIndex) - m_MaxBorder[attrIndex]) / norm;
} else if (inst.value(attrIndex) < m_MinBorder[attrIndex]) {
return (m_MinBorder[attrIndex] - inst.value(attrIndex)) / norm;
} else {
return 0;
}
/* nominal attribute */
} else {
if(holds(attrIndex, inst.value(attrIndex))){
return 0;
} else {
return 1;
}
}
}
/**
* Returns the square of the distance between inst and the Exemplar.
*
* @param inst an instance
* @return the squared distance between inst and the Exemplar.
*/
private double squaredDistance(Instance inst) {
double sum = 0, term;
int numNotMissingAttr = 0;
for(int i = 0; i < inst.numAttributes(); i++){
if(i == classIndex())
continue;
term = m_NNge.attrWeight(i) * attrDistance(inst, i);
term = term * term;
sum += term;
if (!inst.isMissing(i))
numNotMissingAttr++;
}
if(numNotMissingAttr == 0){
return 0;
} else {
return sum / (double) (numNotMissingAttr * numNotMissingAttr);
}
}
/**
* Return the weight of the Examplar
*
* @return the weight of the Examplar.
*/
private double weight(){
return ((double) (m_PositiveCount + m_NegativeCount)) / ((double) m_PositiveCount);
}
/**
* Return the class of the Exemplar
*
* @return the class of this exemplar as a double (weka format)
*/
private double classValue(){
return m_ClassValue;
}
/**
* Returns the value of the inf border of the Exemplar.
*
* @param attrIndex the index of the attribute
* @return the value of the inf border for this attribute
* @throws Exception is thrown either if the attribute is nominal or if the Exemplar is empty
*/
private double getMinBorder(int attrIndex) throws Exception {
if(!attribute(attrIndex).isNumeric())
throw new Exception("Exception.getMinBorder : not numeric attribute !");
if(numInstances() == 0)
throw new Exception("Exception.getMinBorder : empty Exemplar !");
return m_MinBorder[attrIndex];
}
/**
* Returns the value of the sup border of the hyperrectangle
* Returns NaN if the HyperRectangle doesn't have any border for this attribute
*
* @param attrIndex the index of the attribute
* @return the value of the sup border for this attribute
* @throws Exception is thrown either if the attribute is nominal or if the Exemplar is empty
*/
private double getMaxBorder(int attrIndex) throws Exception {
if(!attribute(attrIndex).isNumeric())
throw new Exception("Exception.getMaxBorder : not numeric attribute !");
if(numInstances() == 0)
throw new Exception("Exception.getMaxBorder : empty Exemplar !");
return m_MaxBorder[attrIndex];
}
/**
* Returns the number of positive classifications
*
* @return the number of positive classifications
*/
private int getPositiveCount(){
return m_PositiveCount;
}
/**
* Returns the number of negative classifications
*
* @return the number of negative classifications
*/
private int getNegativeCount(){
return m_NegativeCount;
}
/**
* Set the number of positive classifications
*
* @param value an integer value (greater than 0 is wise...)
*/
private void setPositiveCount(int value) {
m_PositiveCount = value;
}
/**
* Set the number of negative classifications
*
* @param value an integer value
*/
private void setNegativeCount(int value) {
m_NegativeCount = value;
}
/**
* Increment the number of positive Classifications
*/
private void incrPositiveCount(){
m_PositiveCount++;
}
/**
* Increment the number of negative Classifications
*/
private void incrNegativeCount(){
m_NegativeCount++;
}
/**
* Returns true if the Exemplar is empty (i.e. doesn't yield any Instance)
*
* @return true if the Exemplar is empty, false otherwise
*/
public boolean isEmpty(){
return (numInstances() == 0);
}
/**
* Returns a description of this Exemplar
*
* @return A string that describes this Exemplar
*/
private String toString2(){
String s;
Enumeration enu = null;
s = "Exemplar[";
if (numInstances() == 0) {
return s + "Empty]";
}
s += "{";
enu = enumerateInstances();
while(enu.hasMoreElements()){
s = s + "<" + enu.nextElement().toString() + "> ";
}
s = s.substring(0, s.length()-1);
s = s + "} {" + toRules() + "} p=" + m_PositiveCount + " n=" + m_NegativeCount + "]";
return s;
}
/**
* Returns a string of the rules induced by this examplar
*
* @return a string of the rules induced by this examplar
*/
private String toRules(){
if (numInstances() == 0)
return "No Rules (Empty Exemplar)";
String s = "", sep = "";
for(int i = 0; i < numAttributes(); i++){
if(i == classIndex())
continue;
if(attribute(i).isNumeric()){
if(m_MaxBorder[i] != m_MinBorder[i]){
s += sep + m_MinBorder[i] + "<=" + attribute(i).name() + "<=" + m_MaxBorder[i];
} else {
s += sep + attribute(i).name() + "=" + m_MaxBorder[i];
}
sep = " ^ ";
} else {
s += sep + attribute(i).name() + " in {";
String virg = "";
for(int j = 0; j < attribute(i).numValues() + 1; j++){
if(m_Range[i][j]){
s+= virg;
if(j == attribute(i).numValues())
s += "?";
else
s += attribute(i).value(j);
virg = ",";
}
}
s+="}";
sep = " ^ ";
}
}
s += " ("+numInstances() +")";
return s;
}
/**
* Returns the revision string.
*
* @return the revision
*/
public String getRevision() {
return RevisionUtils.extract("$Revision: 5529 $");
}
}
/** An empty instances to keep the headers, the classIndex, etc... */
private Instances m_Train;
/** The list of Exemplars */
private Exemplar m_Exemplars;
/** The lists of Exemplars by class */
private Exemplar m_ExemplarsByClass[];
/** The minimum values for numeric attributes. */
double [] m_MinArray;
/** The maximum values for numeric attributes. */
double [] m_MaxArray;
/** The number of try for generalisation */
private int m_NumAttemptsOfGene = 5;
/** The number of folder for the Mutual Information */
private int m_NumFoldersMI = 5;
/** Values to use for missing value */
private double [] m_MissingVector;
/** MUTUAL INFORMATION'S DATAS */
/* numeric attributes */
private int [][][] m_MI_NumAttrClassInter;
private int [][] m_MI_NumAttrInter;
private double [] m_MI_MaxArray;
private double [] m_MI_MinArray;
/* nominal attributes */
private int [][][] m_MI_NumAttrClassValue;
private int [][] m_MI_NumAttrValue;
/* both */
private int [] m_MI_NumClass;
private int m_MI_NumInst;
private double [] m_MI;
/** MAIN FUNCTIONS OF THE CLASSIFIER */
/**
* Returns default capabilities of the classifier.
*
* @return the capabilities of this classifier
*/
public Capabilities getCapabilities() {
Capabilities result = super.getCapabilities();
result.disableAll();
// attributes
result.enable(Capability.NOMINAL_ATTRIBUTES);
result.enable(Capability.NUMERIC_ATTRIBUTES);
result.enable(Capability.DATE_ATTRIBUTES);
result.enable(Capability.MISSING_VALUES);
// class
result.enable(Capability.NOMINAL_CLASS);
result.enable(Capability.MISSING_CLASS_VALUES);
// instances
result.setMinimumNumberInstances(0);
return result;
}
/**
* Generates a classifier. Must initialize all fields of the classifier
* that are not being set via options (ie. multiple calls of buildClassifier
* must always lead to the same result). Must not change the dataset
* in any way.
*
* @param data set of instances serving as training data
* @throws Exception if the classifier has not been
* generated successfully
*/
public void buildClassifier(Instances data) throws Exception {
// can classifier handle the data?
getCapabilities().testWithFail(data);
// remove instances with missing class
data = new Instances(data);
data.deleteWithMissingClass();
/* initialize the classifier */
m_Train = new Instances(data, 0);
m_Exemplars = null;
m_ExemplarsByClass = new Exemplar[m_Train.numClasses()];
for(int i = 0; i < m_Train.numClasses(); i++){
m_ExemplarsByClass[i] = null;
}
m_MaxArray = new double[m_Train.numAttributes()];
m_MinArray = new double[m_Train.numAttributes()];
for(int i = 0; i < m_Train.numAttributes(); i++){
m_MinArray[i] = Double.POSITIVE_INFINITY;
m_MaxArray[i] = Double.NEGATIVE_INFINITY;
}
m_MI_MinArray = new double [data.numAttributes()];
m_MI_MaxArray = new double [data.numAttributes()];
m_MI_NumAttrClassInter = new int[data.numAttributes()][][];
m_MI_NumAttrInter = new int[data.numAttributes()][];
m_MI_NumAttrClassValue = new int[data.numAttributes()][][];
m_MI_NumAttrValue = new int[data.numAttributes()][];
m_MI_NumClass = new int[data.numClasses()];
m_MI = new double[data.numAttributes()];
m_MI_NumInst = 0;
for(int cclass = 0; cclass < data.numClasses(); cclass++)
m_MI_NumClass[cclass] = 0;
for (int attrIndex = 0; attrIndex < data.numAttributes(); attrIndex++) {
if(attrIndex == data.classIndex())
continue;
m_MI_MaxArray[attrIndex] = m_MI_MinArray[attrIndex] = Double.NaN;
m_MI[attrIndex] = Double.NaN;
if(data.attribute(attrIndex).isNumeric()){
m_MI_NumAttrInter[attrIndex] = new int[m_NumFoldersMI];
for(int inter = 0; inter < m_NumFoldersMI; inter++){
m_MI_NumAttrInter[attrIndex][inter] = 0;
}
} else {
m_MI_NumAttrValue[attrIndex] = new int[data.attribute(attrIndex).numValues() + 1];
for(int attrValue = 0; attrValue < data.attribute(attrIndex).numValues() + 1; attrValue++){
m_MI_NumAttrValue[attrIndex][attrValue] = 0;
}
}
m_MI_NumAttrClassInter[attrIndex] = new int[data.numClasses()][];
m_MI_NumAttrClassValue[attrIndex] = new int[data.numClasses()][];
for(int cclass = 0; cclass < data.numClasses(); cclass++){
if(data.attribute(attrIndex).isNumeric()){
m_MI_NumAttrClassInter[attrIndex][cclass] = new int[m_NumFoldersMI];
for(int inter = 0; inter < m_NumFoldersMI; inter++){
m_MI_NumAttrClassInter[attrIndex][cclass][inter] = 0;
}
} else if(data.attribute(attrIndex).isNominal()){
m_MI_NumAttrClassValue[attrIndex][cclass] = new int[data.attribute(attrIndex).numValues() + 1];
for(int attrValue = 0; attrValue < data.attribute(attrIndex).numValues() + 1; attrValue++){
m_MI_NumAttrClassValue[attrIndex][cclass][attrValue] = 0;
}
}
}
}
m_MissingVector = new double[data.numAttributes()];
for(int i = 0; i < data.numAttributes(); i++){
if(i == data.classIndex()){
m_MissingVector[i] = Double.NaN;
} else {
m_MissingVector[i] = data.attribute(i).numValues();
}
}
/* update the classifier with data */
Enumeration enu = data.enumerateInstances();
while(enu.hasMoreElements()){
update((Instance) enu.nextElement());
}
}
/**
* Classifies a given instance.
*
* @param instance the instance to be classified
* @return index of the predicted class as a double
* @throws Exception if instance could not be classified
* successfully
*/
public double classifyInstance(Instance instance) throws Exception {
/* check the instance */
if (m_Train.equalHeaders(instance.dataset()) == false){
throw new Exception("NNge.classifyInstance : Incompatible instance types !");
}
Exemplar matched = nearestExemplar(instance);
if(matched == null){
throw new Exception("NNge.classifyInstance : NNge hasn't been trained !");
}
return matched.classValue();
}
/**
* Updates the classifier using the given instance.
*
* @param instance the instance to include
* @throws Exception if instance could not be incorporated
* successfully
*/
public void updateClassifier(Instance instance) throws Exception {
if (m_Train.equalHeaders(instance.dataset()) == false) {
throw new Exception("Incompatible instance types");
}
update(instance);
}
/** HIGH LEVEL SUB-FUNCTIONS */
/**
* Performs the update of the classifier
*
* @param instance the new instance
* @throws Exception if the update fails
*/
private void update(Instance instance) throws Exception {
if (instance.classIsMissing()) {
return;
}
instance.replaceMissingValues(m_MissingVector);
m_Train.add(instance);
/* Update the minimum and maximum for all the attributes */
updateMinMax(instance);
/* update the mutual information datas */
updateMI(instance);
/* Nearest Exemplar */
Exemplar nearest = nearestExemplar(instance);
/* Adjust */
if(nearest == null){
Exemplar newEx = new Exemplar(this, m_Train, 10, instance.classValue());
newEx.generalise(instance);
initWeight(newEx);
addExemplar(newEx);
return;
}
adjust(instance, nearest);
/* Generalise */
generalise(instance);
}
/**
* Returns the nearest Exemplar
*
* @param inst an Instance
* @return the nearest Exemplar to inst, null if no exemplar are found.
*/
private Exemplar nearestExemplar(Instance inst){
if (m_Exemplars == null)
return null;
Exemplar cur = m_Exemplars, nearest = m_Exemplars;
double dist, smallestDist = cur.squaredDistance(inst);
while (cur.next != null){
cur = cur.next;
dist = cur.squaredDistance(inst);
if (dist < smallestDist){
smallestDist = dist;
nearest = cur;
}
}
return nearest;
}
/**
* Returns the nearest Exemplar with class c
*
* @param inst an Instance
* @param c the class of the Exemplar to return
* @return the nearest Exemplar to inst with class c, null if no exemplar with class c are found.
*/
private Exemplar nearestExemplar(Instance inst, double c){
if (m_ExemplarsByClass[(int) c] == null)
return null;
Exemplar cur = m_ExemplarsByClass[(int) c], nearest = m_ExemplarsByClass[(int) c];
double dist, smallestDist = cur.squaredDistance(inst);
while (cur.nextWithClass != null){
cur = cur.nextWithClass;
dist = cur.squaredDistance(inst);
if (dist < smallestDist){
smallestDist = dist;
nearest = cur;
}
}
return nearest;
}
/**
* Generalise an Exemplar (not necessarily predictedExemplar) to match instance.
* predictedExemplar must be in NNge's lists
*
* @param newInst the new instance
* @throws Exception in case of inconsitent situation
*/
private void generalise(Instance newInst) throws Exception {
Exemplar first = m_ExemplarsByClass[(int) newInst.classValue()];
int n = 0;
/* try to generalise with the n first exemplars */
while(n < m_NumAttemptsOfGene && first != null){
/* find the nearest one starting from first */
Exemplar closest = first, cur = first;
double smallestDist = first.squaredDistance(newInst), dist;
while(cur.nextWithClass != null){
cur = cur.nextWithClass;
dist = cur.squaredDistance(newInst);
if(dist < smallestDist){
smallestDist = dist;
closest = cur;
}
}
/* remove the Examplar from NNge's lists */
if(closest == first)
first = first.nextWithClass;
removeExemplar(closest);
/* try to generalise */
closest.preGeneralise(newInst);
if(!detectOverlapping(closest)){
closest.validateGeneralisation();
addExemplar(closest);
return;
}
/* it didn't work, put ungeneralised exemplar on the top of the lists */
closest.cancelGeneralisation();
addExemplar(closest);
n++;
}
/* generalisation failled : add newInst as a new Examplar */
Exemplar newEx = new Exemplar(this, m_Train, 5, newInst.classValue());
newEx.generalise(newInst);
initWeight(newEx);
addExemplar(newEx);
}
/**
* Adjust the NNge.
*
* @param newInst the instance to classify
* @param predictedExemplar the Exemplar that matches newInst
* @throws Exception in case of inconsistent situation
*/
private void adjust(Instance newInst, Exemplar predictedExemplar) throws Exception {
/* correct prediction */
if(newInst.classValue() == predictedExemplar.classValue()){
predictedExemplar.incrPositiveCount();
/* incorrect prediction */
} else {
predictedExemplar.incrNegativeCount();
/* new instance falls inside */
if(predictedExemplar.holds(newInst)){
prune(predictedExemplar, newInst);
}
}
}
/**
* Prunes an Exemplar that matches an Instance
*
* @param predictedExemplar an Exemplar
* @param newInst an Instance matched by predictedExemplar
* @throws Exception in case of inconsistent situation. (shouldn't happen.)
*/
private void prune(Exemplar predictedExemplar, Instance newInst) throws Exception {
/* remove the Exemplar */
removeExemplar(predictedExemplar);
/* look for the best nominal feature and the best numeric feature to cut */
int numAttr = -1, nomAttr = -1;
double smallestDelta = Double.POSITIVE_INFINITY, delta;
int biggest_N_Nom = -1, biggest_N_Num = -1, n, m;
for(int i = 0; i < m_Train.numAttributes(); i++){
if(i == m_Train.classIndex())
continue;
/* numeric attribute */
if(m_Train.attribute(i).isNumeric()){
/* compute the distance 'delta' to the closest boundary */
double norm = m_MaxArray[i] - m_MinArray[i];
if(norm != 0){
delta = Math.min((predictedExemplar.getMaxBorder(i) - newInst.value(i)),
(newInst.value(i) - predictedExemplar.getMinBorder(i))) / norm;
} else {
delta = Double.POSITIVE_INFINITY;
}
/* compute the size of the biggest Exemplar which would be created */
n = m = 0;
Enumeration enu = predictedExemplar.enumerateInstances();
while(enu.hasMoreElements()){
Instance ins = (Instance) enu.nextElement();
if(ins.value(i) < newInst.value(i))
n++;
else if(ins.value(i) > newInst.value(i))
m++;
}
n = Math.max(n, m);
if(delta < smallestDelta){
smallestDelta = delta;
biggest_N_Num = n;
numAttr = i;
} else if(delta == smallestDelta && n > biggest_N_Num){
biggest_N_Num = n;
numAttr = i;
}
/* nominal attribute */
} else {
/* compute the size of the Exemplar which would be created */
Enumeration enu = predictedExemplar.enumerateInstances();
n = 0;
while(enu.hasMoreElements()){
if(((Instance) enu.nextElement()).value(i) != newInst.value(i))
n++;
}
if(n > biggest_N_Nom){
biggest_N_Nom = n;
nomAttr = i;
}
}
}
/* selection of the feature to cut between the best nominal and the best numeric */
int attrToCut;
if(numAttr == -1 && nomAttr == -1){
attrToCut = 0;
} else if (numAttr == -1){
attrToCut = nomAttr;
} else if(nomAttr == -1){
attrToCut = numAttr;
} else {
if(biggest_N_Nom > biggest_N_Num)
attrToCut = nomAttr;
else
attrToCut = numAttr;
}
/* split the Exemplar */
Instance curInst;
Exemplar a, b;
a = new Exemplar(this, m_Train, 10, predictedExemplar.classValue());
b = new Exemplar(this, m_Train, 10, predictedExemplar.classValue());
LinkedList leftAlone = new LinkedList();
Enumeration enu = predictedExemplar.enumerateInstances();
if(m_Train.attribute(attrToCut).isNumeric()){
while(enu.hasMoreElements()){
curInst = (Instance) enu.nextElement();
if(curInst.value(attrToCut) > newInst.value(attrToCut)){
a.generalise(curInst);
} else if (curInst.value(attrToCut) < newInst.value(attrToCut)){
b.generalise(curInst);
} else if (notEqualFeatures(curInst, newInst)) {
leftAlone.add(curInst);
}
}
} else {
while(enu.hasMoreElements()){
curInst = (Instance) enu.nextElement();
if(curInst.value(attrToCut) != newInst.value(attrToCut)){
a.generalise(curInst);
} else if (notEqualFeatures(curInst, newInst)){
leftAlone.add(curInst);
}
}
}
/* treat the left alone Instances */
while(leftAlone.size() != 0){
Instance alone = (Instance) leftAlone.removeFirst();
a.preGeneralise(alone);
if(!a.holds(newInst)){
a.validateGeneralisation();
continue;
}
a.cancelGeneralisation();
b.preGeneralise(alone);
if(!b.holds(newInst)){
b.validateGeneralisation();
continue;
}
b.cancelGeneralisation();
Exemplar exem = new Exemplar(this, m_Train, 3, alone.classValue());
exem.generalise(alone);
initWeight(exem);
addExemplar(exem);
}
/* add (or not) the new Exemplars */
if(a.numInstances() != 0){
initWeight(a);
addExemplar(a);
}
if(b.numInstances() != 0){
initWeight(b);
addExemplar(b);
}
}
/**
* Returns true if the instance don't have the same feature values
*
* @param inst1 an instance
* @param inst2 an instance
* @return true if the instance don't have the same feature values
*/
private boolean notEqualFeatures(Instance inst1, Instance inst2) {
for(int i = 0; i < m_Train.numAttributes(); i++){
if(i == m_Train.classIndex())
continue;
if(inst1.value(i) != inst2.value(i))
return true;
}
return false;
}
/**
* Returns true if ex overlaps any of the Exemplars in NNge's lists
*
* @param ex an Exemplars
* @return true if ex overlaps any of the Exemplars in NNge's lists
*/
private boolean detectOverlapping(Exemplar ex){
Exemplar cur = m_Exemplars;
while(cur != null){
if(ex.overlaps(cur)){
return true;
}
cur = cur.next;
}
return false;
}
/**
* Updates the minimum, maximum, sum, sumSquare values for all the attributes
*
* @param instance the new instance
*/
private void updateMinMax(Instance instance){
for (int j = 0; j < m_Train.numAttributes(); j++) {
if(m_Train.classIndex() == j || m_Train.attribute(j).isNominal())
continue;
if (instance.value(j) < m_MinArray[j])
m_MinArray[j] = instance.value(j);
if (instance.value(j) > m_MaxArray[j])
m_MaxArray[j] = instance.value(j);
}
}
/**
* Updates the data for computing the mutual information
*
* MUST be called AFTER adding inst in m_Train
*
* @param inst the new instance
* @throws Exception is thrown if an inconsistent situation is met
*/
private void updateMI(Instance inst) throws Exception {
if(m_NumFoldersMI < 1){
throw new Exception("NNge.updateMI : incorrect number of folders ! Option I must be greater than 1.");
}
m_MI_NumClass[(int) inst.classValue()]++;
m_MI_NumInst++;
/* for each attribute */
for(int attrIndex = 0; attrIndex < m_Train.numAttributes(); attrIndex++){
/* which is the class attribute */
if(m_Train.classIndex() == attrIndex)
continue;
/* which is a numeric attribute */
else if(m_Train.attribute(attrIndex).isNumeric()){
/* if max-min have to be updated */
if(Double.isNaN(m_MI_MaxArray[attrIndex]) ||
Double.isNaN(m_MI_MinArray[attrIndex]) ||
m_MI_MaxArray[attrIndex] < inst.value(attrIndex) ||
inst.value(attrIndex) < m_MI_MinArray[attrIndex]){
/* then update them */
if(Double.isNaN(m_MI_MaxArray[attrIndex])) m_MI_MaxArray[attrIndex] = inst.value(attrIndex);
if(Double.isNaN(m_MI_MinArray[attrIndex])) m_MI_MinArray[attrIndex] = inst.value(attrIndex);
if(m_MI_MaxArray[attrIndex] < inst.value(attrIndex)) m_MI_MaxArray[attrIndex] = inst.value(attrIndex);
if(m_MI_MinArray[attrIndex] > inst.value(attrIndex)) m_MI_MinArray[attrIndex] = inst.value(attrIndex);
/* and re-compute everything from scratch... (just for this attribute) */
double delta = (m_MI_MaxArray[attrIndex] - m_MI_MinArray[attrIndex]) / (double) m_NumFoldersMI;
/* for each interval */
for(int inter = 0; inter < m_NumFoldersMI; inter++){
m_MI_NumAttrInter[attrIndex][inter] = 0;
/* for each class */
for(int cclass = 0; cclass < m_Train.numClasses(); cclass++){
m_MI_NumAttrClassInter[attrIndex][cclass][inter] = 0;
/* count */
Enumeration enu = m_Train.enumerateInstances();
while(enu.hasMoreElements()){
Instance cur = (Instance) enu.nextElement();
if(( (m_MI_MinArray[attrIndex] + inter * delta) <= cur.value(attrIndex) ) &&
( cur.value(attrIndex) <= (m_MI_MinArray[attrIndex] + (inter + 1) * delta) ) &&
( cur.classValue() == cclass ) ){
m_MI_NumAttrInter[attrIndex][inter]++;
m_MI_NumAttrClassInter[attrIndex][cclass][inter]++;
}
}
}
}
/* max-min don't have to be updated */
} else {
/* still have to incr the card of the correct interval */
double delta = (m_MI_MaxArray[attrIndex] - m_MI_MinArray[attrIndex]) / (double) m_NumFoldersMI;
/* for each interval */
for(int inter = 0; inter < m_NumFoldersMI; inter++){
/* which contains inst*/
if(( (m_MI_MinArray[attrIndex] + inter * delta) <= inst.value(attrIndex) ) &&
( inst.value(attrIndex) <= (m_MI_MinArray[attrIndex] + (inter + 1) * delta) )){
m_MI_NumAttrInter[attrIndex][inter]++;
m_MI_NumAttrClassInter[attrIndex][(int) inst.classValue()][inter]++;
}
}
}
/* update the mutual information of this attribute... */
m_MI[attrIndex] = 0;
/* for each interval, for each class */
for(int inter = 0; inter < m_NumFoldersMI; inter++){
for(int cclass = 0; cclass < m_Train.numClasses(); cclass++){
double pXY = ((double) m_MI_NumAttrClassInter[attrIndex][cclass][inter]) / ((double) m_MI_NumInst);
double pX = ((double) m_MI_NumClass[cclass]) / ((double) m_MI_NumInst);
double pY = ((double) m_MI_NumAttrInter[attrIndex][inter]) / ((double) m_MI_NumInst);
if(pXY != 0)
m_MI[attrIndex] += pXY * Utils.log2(pXY / (pX * pY));
}
}
/* which is a nominal attribute */
} else if (m_Train.attribute(attrIndex).isNominal()){
/*incr the card of the correct 'values' */
m_MI_NumAttrValue[attrIndex][(int) inst.value(attrIndex)]++;
m_MI_NumAttrClassValue[attrIndex][(int) inst.classValue()][(int) inst.value(attrIndex)]++;
/* update the mutual information of this attribute... */
m_MI[attrIndex] = 0;
/* for each nominal value, for each class */
for(int attrValue = 0; attrValue < m_Train.attribute(attrIndex).numValues() + 1; attrValue++){
for(int cclass = 0; cclass < m_Train.numClasses(); cclass++){
double pXY = ((double) m_MI_NumAttrClassValue[attrIndex][cclass][attrValue]) / ((double) m_MI_NumInst);
double pX = ((double) m_MI_NumClass[cclass]) / ((double) m_MI_NumInst);
double pY = ((double) m_MI_NumAttrValue[attrIndex][attrValue]) / ((double) m_MI_NumInst);
if(pXY != 0)
m_MI[attrIndex] += pXY * Utils.log2(pXY / (pX * pY));
}
}
/* not a nominal attribute, not a numeric attribute */
} else {
throw new Exception("NNge.updateMI : Cannot deal with 'string attribute'.");
}
}
}
/**
* Init the weight of ex
* Watch out ! ex shouldn't be in NNge's lists when initialized
*
* @param ex the Exemplar to initialise
*/
private void initWeight(Exemplar ex) {
int pos = 0, neg = 0, n = 0;
Exemplar cur = m_Exemplars;
if (cur == null){
ex.setPositiveCount(1);
ex.setNegativeCount(0);
return;
}
while(cur != null){
pos += cur.getPositiveCount();
neg += cur.getNegativeCount();
n++;
cur = cur.next;
}
ex.setPositiveCount(pos / n);
ex.setNegativeCount(neg / n);
}
/**
* Adds an Exemplar in NNge's lists
* Ensure that the exemplar is not already in a list : the links would be broken...
*
* @param ex a new Exemplar to add
*/
private void addExemplar(Exemplar ex) {
/* add ex at the top of the general list */
ex.next = m_Exemplars;
if(m_Exemplars != null)
m_Exemplars.previous = ex;
ex.previous = null;
m_Exemplars = ex;
/* add ex at the top of the corresponding class list */
ex.nextWithClass = m_ExemplarsByClass[(int) ex.classValue()];
if(m_ExemplarsByClass[(int) ex.classValue()] != null)
m_ExemplarsByClass[(int) ex.classValue()].previousWithClass = ex;
ex.previousWithClass = null;
m_ExemplarsByClass[(int) ex.classValue()] = ex;
}
/**
* Removes an Exemplar from NNge's lists
* Ensure that the Exemplar is actually in NNge's lists.
* Likely to do something wrong if this condition is not respected.
* Due to the list implementation, the Exemplar can appear only once in the lists :
* once removed, the exemplar is not in the lists anymore.
*
* @param ex a new Exemplar to add
*/
private void removeExemplar(Exemplar ex){
/* remove from the general list */
if(m_Exemplars == ex){
m_Exemplars = ex.next;
if(m_Exemplars != null)
m_Exemplars.previous = null;
} else {
ex.previous.next = ex.next;
if(ex.next != null){
ex.next.previous = ex.previous;
}
}
ex.next = ex.previous = null;
/* remove from the class list */
if(m_ExemplarsByClass[(int) ex.classValue()] == ex){
m_ExemplarsByClass[(int) ex.classValue()] = ex.nextWithClass;
if(m_ExemplarsByClass[(int) ex.classValue()] != null)
m_ExemplarsByClass[(int) ex.classValue()].previousWithClass = null;
} else {
ex.previousWithClass.nextWithClass = ex.nextWithClass;
if(ex.nextWithClass != null){
ex.nextWithClass.previousWithClass = ex.previousWithClass;
}
}
ex.nextWithClass = ex.previousWithClass = null;
}
/**
* returns the weight of indexth attribute
*
* @param index attribute's index
* @return the weight of indexth attribute
*/
private double attrWeight (int index) {
return m_MI[index];
}
/**
* Returns a description of this classifier.
*
* @return a description of this classifier as a string.
*/
public String toString(){
String s;
Exemplar cur = m_Exemplars;
int i;
if (m_MinArray == null) {
return "No classifier built";
}
int[] nbHypClass = new int[m_Train.numClasses()];
int[] nbSingleClass = new int[m_Train.numClasses()];
for(i = 0; i<nbHypClass.length; i++){
nbHypClass[i] = 0;
nbSingleClass[i] = 0;
}
int nbHyp = 0, nbSingle = 0;
s = "\nNNGE classifier\n\nRules generated :\n";
while(cur != null){
s += "\tclass " + m_Train.attribute(m_Train.classIndex()).value((int) cur.classValue()) + " IF : ";
s += cur.toRules() + "\n";
nbHyp++;
nbHypClass[(int) cur.classValue()]++;
if (cur.numInstances() == 1){
nbSingle++;
nbSingleClass[(int) cur.classValue()]++;
}
cur = cur.next;
}
s += "\nStat :\n";
for(i = 0; i<nbHypClass.length; i++){
s += "\tclass " + m_Train.attribute(m_Train.classIndex()).value(i) +
" : " + Integer.toString(nbHypClass[i]) + " exemplar(s) including " +
Integer.toString(nbHypClass[i] - nbSingleClass[i]) + " Hyperrectangle(s) and " +
Integer.toString(nbSingleClass[i]) + " Single(s).\n";
}
s += "\n\tTotal : " + Integer.toString(nbHyp) + " exemplars(s) including " +
Integer.toString(nbHyp - nbSingle) + " Hyperrectangle(s) and " +
Integer.toString(nbSingle) + " Single(s).\n";
s += "\n";
s += "\tFeature weights : ";
String space = "[";
for(int ii = 0; ii < m_Train.numAttributes(); ii++){
if(ii != m_Train.classIndex()){
s += space + Double.toString(attrWeight(ii));
space = " ";
}
}
s += "]";
s += "\n\n";
return s;
}
/** OPTION HANDLER FUNCTION */
/**
* Returns an enumeration of all the available options..
*
* @return an enumeration of all available options.
*/
public Enumeration listOptions(){
Vector newVector = new Vector(2);
newVector.addElement(new Option(
"\tNumber of attempts of generalisation.\n",
"G",
1,
"-G <value>"));
newVector.addElement(new Option(
"\tNumber of folder for computing the mutual information.\n",
"I",
1,
"-I <value>"));
return newVector.elements();
}
/**
* Sets the OptionHandler's options using the given list. All options
* will be set (or reset) during this call (i.e. incremental setting
* of options is not possible). <p/>
*
<!-- options-start -->
* Valid options are: <p/>
*
* <pre> -G <value>
* Number of attempts of generalisation.
* </pre>
*
* <pre> -I <value>
* Number of folder for computing the mutual information.
* </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 {
String str;
/* Number max of attempts of generalisation */
str = Utils.getOption('G', options);
if(str.length() != 0){
m_NumAttemptsOfGene = Integer.parseInt(str);
if(m_NumAttemptsOfGene < 1)
throw new Exception("NNge.setOptions : G option's value must be greater than 1.");
} else {
m_NumAttemptsOfGene = 5;
}
/* Number of folder for computing the mutual information */
str = Utils.getOption('I', options);
if(str.length() != 0){
m_NumFoldersMI = Integer.parseInt(str);
if(m_NumFoldersMI < 1)
throw new Exception("NNge.setOptions : I option's value must be greater than 1.");
} else {
m_NumFoldersMI = 5;
}
}
/**
* Gets the current option settings for the OptionHandler.
*
* @return the list of current option settings as an array of strings
*/
public String[] getOptions(){
String[] options = new String[5];
int current = 0;
options[current++] = "-G"; options[current++] = "" + m_NumAttemptsOfGene;
options[current++] = "-I"; options[current++] = "" + m_NumFoldersMI;
while (current < options.length) {
options[current++] = "";
}
return options;
}
/**
* Returns the tip text for this property
* @return tip text for this property suitable for
* displaying in the explorer/experimenter gui
*/
public String numAttemptsOfGeneOptionTipText() {
return "Sets the number of attempts for generalization.";
}
/**
* Gets the number of attempts for generalisation.
*
* @return the value of the option G
*/
public int getNumAttemptsOfGeneOption() {
return m_NumAttemptsOfGene;
}
/**
* Sets the number of attempts for generalisation.
*
* @param newIntParameter the new value.
*/
public void setNumAttemptsOfGeneOption(int newIntParameter) {
m_NumAttemptsOfGene = newIntParameter;
}
/**
* Returns the tip text for this property
* @return tip text for this property suitable for
* displaying in the explorer/experimenter gui
*/
public String numFoldersMIOptionTipText() {
return "Sets the number of folder for mutual information.";
}
/**
* Gets the number of folder for mutual information.
*
* @return the value of the option I
*/
public int getNumFoldersMIOption() {
return m_NumFoldersMI;
}
/**
* Sets the number of folder for mutual information.
*
* @param newIntParameter the new value.
*/
public void setNumFoldersMIOption(int newIntParameter) {
m_NumFoldersMI = newIntParameter;
}
/**
* Returns the revision string.
*
* @return the revision
*/
public String getRevision() {
return RevisionUtils.extract("$Revision: 5529 $");
}
/**
* Main method for testing this class.
*
* @param argv should contain command line arguments for evaluation
* (see Evaluation).
*/
public static void main(String [] argv) {
runClassifier(new NNge(), argv);
}
}