[Home] DMelt member access . Java source code: weka.filters.unsupervised.attribute.RandomProjection
/*
* 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 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 General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see .
*/
/*
* RandomProjection.java
* Copyright (C) 2003-2012 University of Waikato, Hamilton, New Zealand
*
*/
package weka.filters.unsupervised.attribute;
import java.util.ArrayList;
import java.util.Enumeration;
import java.util.Random;
import java.util.Vector;
import weka.core.*;
import weka.core.Capabilities.Capability;
import weka.core.TechnicalInformation.Field;
import weka.core.TechnicalInformation.Type;
import weka.filters.Filter;
import weka.filters.UnsupervisedFilter;
/**
*
* Reduces the dimensionality of the data by projecting it onto a lower dimensional subspace using a random matrix with columns of unit length. It will reduce the number of attributes in the data while preserving much of its variation like PCA, but at a much less computational cost.
* It first applies the NominalToBinary filter to convert all attributes to numeric before reducing the dimension. It preserves the class attribute.
*
* For more information, see:
*
* Dmitriy Fradkin, David Madigan: Experiments with random projections for machine learning. In: KDD '03: Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, New York, NY, USA, 517-522, 003.
*
*
*
*
* BibTeX:
*
* @inproceedings{Fradkin003,
* address = {New York, NY, USA},
* author = {Dmitriy Fradkin and David Madigan},
* booktitle = {KDD '03: Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining},
* pages = {517-522},
* publisher = {ACM Press},
* title = {Experiments with random projections for machine learning},
* year = {003}
* }
*
*
*
*
*
* Valid options are:
*
* -N <number>
* The number of dimensions (attributes) the data should be reduced to
* (default 10; exclusive of the class attribute, if it is set).
*
* -D [SPARSE1|SPARSE2|GAUSSIAN]
* The distribution to use for calculating the random matrix.
* Sparse1 is:
* sqrt(3)*{-1 with prob(1/6), 0 with prob(2/3), +1 with prob(1/6)}
* Sparse2 is:
* {-1 with prob(1/2), +1 with prob(1/2)}
*
*
* -P <percent>
* The percentage of dimensions (attributes) the data should
* be reduced to (exclusive of the class attribute, if it is set). The -N
* option is ignored if this option is present and is greater
* than zero.
*
* -M
* Replace missing values using the ReplaceMissingValues filter instead of just skipping them.
*
* -R <num>
* The random seed for the random number generator used for
* calculating the random matrix (default 42).
*
*
*
* @author Ashraf M. Kibriya (amk14@cs.waikato.ac.nz)
* @version $Revision: 14609 $ [1.0 - 22 July 2003 - Initial version (Ashraf M.
* Kibriya)]
*/
public class RandomProjection extends Filter implements UnsupervisedFilter,
OptionHandler, TechnicalInformationHandler, Randomizable, WeightedInstancesHandler {
/** for serialization */
static final long serialVersionUID = 4428905532728645880L;
/** Stores the number of dimensions to reduce the data to */
protected int m_k = 10;
/**
* Stores the dimensionality the data should be reduced to as percentage of
* the original dimension
*/
protected double m_percent = 0.0;
/** distribution type: sparse 1 */
public static final int SPARSE1 = 1;
/** distribution type: sparse 2 */
public static final int SPARSE2 = 2;
/** distribution type: gaussian */
public static final int GAUSSIAN = 3;
/**
* The types of distributions that can be used for calculating the random
* matrix
*/
public static final Tag[] TAGS_DSTRS_TYPE = { new Tag(SPARSE1, "Sparse1"),
new Tag(SPARSE2, "Sparse2"), new Tag(GAUSSIAN, "Gaussian"), };
/**
* Stores the distribution to use for calculating the random matrix
*/
protected int m_distribution = SPARSE1;
/**
* Should the missing values be replaced using
* unsupervised.ReplaceMissingValues filter
*/
protected boolean m_useReplaceMissing = false;
/** Keeps track of output format if it is defined or not */
protected boolean m_OutputFormatDefined = false;
/** The NominalToBinary filter applied to the data before this filter */
protected Filter m_ntob;
/** The ReplaceMissingValues filter */
protected Filter m_replaceMissing;
/** Stores the random seed used to generate the random matrix */
protected int m_rndmSeed = 42;
/** The random matrix */
protected double m_rmatrix[][];
/** The random number generator used for generating the random matrix */
protected Random m_random;
/**
* Returns an enumeration describing the available options.
*
* @return an enumeration of all the available options.
*/
@Override
public Enumeration listOptions() {
Vector newVector = new Vector (5);
newVector.addElement(new Option(
"\tThe number of dimensions (attributes) the data should be reduced to\n"
+ "\t(default 10; exclusive of the class attribute, if it is set).",
"N", 1, "-N "));
newVector
.addElement(new Option(
"\tThe distribution to use for calculating the random matrix.\n"
+ "\tSparse1 is:\n"
+ "\t sqrt(3)*{-1 with prob(1/6), 0 with prob(2/3), +1 with prob(1/6)}\n"
+ "\tSparse2 is:\n" + "\t {-1 with prob(1/2), +1 with prob(1/2)}",
"D", 1, "-D [SPARSE1|SPARSE2|GAUSSIAN]"));
// newVector.addElement(new Option(
// "\tUse Gaussian distribution for calculating the random matrix.",
// "G", 0, "-G"));
newVector
.addElement(new Option(
"\tThe percentage of dimensions (attributes) the data should\n"
+ "\tbe reduced to (exclusive of the class attribute, if it is set). The -N\n"
+ "\toption is ignored if this option is present and is greater\n"
+ "\tthan zero.", "P", 1, "-P "));
newVector.addElement(new Option(
"\tReplace missing values using the ReplaceMissingValues filter instead of just skipping them.",
"M", 0,
"-M"));
newVector.addElement(new Option(
"\tThe random seed for the random number generator used for\n"
+ "\tcalculating the random matrix (default 42).", "R", 0, "-R "));
return newVector.elements();
}
/**
* Parses a given list of options.
*
*
*
* Valid options are:
*
* -N <number>
* The number of dimensions (attributes) the data should be reduced to
* (default 10; exclusive of the class attribute, if it is set).
*
* -D [SPARSE1|SPARSE2|GAUSSIAN]
* The distribution to use for calculating the random matrix.
* Sparse1 is:
* sqrt(3)*{-1 with prob(1/6), 0 with prob(2/3), +1 with prob(1/6)}
* Sparse2 is:
* {-1 with prob(1/2), +1 with prob(1/2)}
*
*
* -P <percent>
* The percentage of dimensions (attributes) the data should
* be reduced to (exclusive of the class attribute, if it is set). The -N
* option is ignored if this option is present and is greater
* than zero.
*
* -M
* Replace missing values using the ReplaceMissingValues filter instead of just skipping them.
*
* -R <num>
* The random seed for the random number generator used for
* calculating the random matrix (default 42).
*
*
*
* @param options the list of options as an array of strings
* @throws Exception if an option is not supported
*/
@Override
public void setOptions(String[] options) throws Exception {
String mString = Utils.getOption('P', options);
if (mString.length() != 0) {
setPercent(Double.parseDouble(mString));
} else {
setPercent(0);
mString = Utils.getOption('N', options);
if (mString.length() != 0) {
setNumberOfAttributes(Integer.parseInt(mString));
} else {
setNumberOfAttributes(10);
}
}
mString = Utils.getOption('R', options);
if (mString.length() != 0) {
setSeed(Integer.parseInt(mString));
}
mString = Utils.getOption('D', options);
if (mString.length() != 0) {
if (mString.equalsIgnoreCase("sparse1")) {
setDistribution(new SelectedTag(SPARSE1, TAGS_DSTRS_TYPE));
} else if (mString.equalsIgnoreCase("sparse2")) {
setDistribution(new SelectedTag(SPARSE2, TAGS_DSTRS_TYPE));
} else if (mString.equalsIgnoreCase("gaussian")) {
setDistribution(new SelectedTag(GAUSSIAN, TAGS_DSTRS_TYPE));
}
}
if (Utils.getFlag('M', options)) {
setReplaceMissingValues(true);
} else {
setReplaceMissingValues(false);
}
Utils.checkForRemainingOptions(options);
}
/**
* Gets the current settings of the filter.
*
* @return an array of strings suitable for passing to setOptions
*/
@Override
public String[] getOptions() {
Vector options = new Vector();
if (getReplaceMissingValues()) {
options.add("-M");
}
if (getPercent() <= 0) {
options.add("-N");
options.add("" + getNumberOfAttributes());
} else {
options.add("-P");
options.add("" + getPercent());
}
options.add("-R");
options.add("" + getSeed());
SelectedTag t = getDistribution();
options.add("-D");
options.add("" + t.getSelectedTag().getReadable());
return options.toArray(new String[0]);
}
/**
* Returns a string describing this filter
*
* @return a description of the filter suitable for displaying in the
* explorer/experimenter gui
*/
public String globalInfo() {
return "Reduces the dimensionality of the data by projecting"
+ " it onto a lower dimensional subspace using a random"
+ " matrix with columns of unit length. It will reduce"
+ " the number of attributes in the data while preserving"
+ " much of its variation like PCA, but at a much less"
+ " computational cost.\n"
+ "It first applies the NominalToBinary filter to"
+ " convert all attributes to numeric before reducing the"
+ " dimension. It preserves the class attribute.\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
*/
@Override
public TechnicalInformation getTechnicalInformation() {
TechnicalInformation result;
result = new TechnicalInformation(Type.INPROCEEDINGS);
result.setValue(Field.AUTHOR, "Dmitriy Fradkin and David Madigan");
result.setValue(Field.TITLE,
"Experiments with random projections for machine learning");
result
.setValue(
Field.BOOKTITLE,
"KDD '03: Proceedings of the ninth ACM SIGKDD International Conference on Knowledge Discovery and Data mining");
result.setValue(Field.YEAR, "003");
result.setValue(Field.PAGES, "517-522");
result.setValue(Field.PUBLISHER, "ACM Press");
result.setValue(Field.ADDRESS, "New York, NY, USA");
return result;
}
/**
* Returns the tip text for this property
*
* @return tip text for this property suitable for displaying in the
* explorer/experimenter gui
*/
public String numberOfAttributesTipText() {
return "The number of dimensions (attributes) the data should"
+ " be reduced to.";
}
/**
* Sets the number of attributes (dimensions) the data should be reduced to
*
* @param newAttNum the goal for the dimensions
*/
public void setNumberOfAttributes(int newAttNum) {
m_k = newAttNum;
}
/**
* Gets the current number of attributes (dimensionality) to which the data
* will be reduced to.
*
* @return the number of dimensions
*/
public int getNumberOfAttributes() {
return m_k;
}
/**
* Returns the tip text for this property
*
* @return tip text for this property suitable for displaying in the
* explorer/experimenter gui
*/
public String percentTipText() {
return " The percentage of dimensions (attributes) the data should"
+ " be reduced to (inclusive of the class attribute). The "
+ " NumberOfAttributes option is ignored if this option is"
+ " present or is greater than zero.";
}
/**
* Sets the percent the attributes (dimensions) of the data should be reduced
* to
*
* @param newPercent the percentage of attributes
*/
public void setPercent(double newPercent) {
if (newPercent > 0) {
newPercent /= 100;
}
m_percent = newPercent;
}
/**
* Gets the percent the attributes (dimensions) of the data will be reduced to
*
* @return the percentage of attributes
*/
public double getPercent() {
return m_percent * 100;
}
/**
* Returns the tip text for this property
*
* @return tip text for this property suitable for displaying in the
* explorer/experimenter gui
*/
public String seedTipText() {
return "The random seed used by the random"
+ " number generator used for generating the random matrix ";
}
/**
* Sets the random seed of the random number generator
*
* @param seed the random seed value
*/
@Override
public void setSeed(int seed) {
m_rndmSeed = seed;
}
/**
* Gets the random seed of the random number generator
*
* @return the random seed value
*/
@Override
public int getSeed() {
return m_rndmSeed;
}
/**
* Returns the tip text for this property
*
* @return tip text for this property suitable for displaying in the
* explorer/experimenter gui
*/
public String distributionTipText() {
return "The distribution to use for calculating the random matrix.\n"
+ "Sparse1 is:\n" + " sqrt(3) * { -1 with prob(1/6), \n"
+ " 0 with prob(2/3), \n"
+ " +1 with prob(1/6) } \n" + "Sparse2 is:\n"
+ " { -1 with prob(1/2), \n" + " +1 with prob(1/2) } ";
}
/**
* Sets the distribution to use for calculating the random matrix
*
* @param newDstr the distribution to use
*/
public void setDistribution(SelectedTag newDstr) {
if (newDstr.getTags() == TAGS_DSTRS_TYPE) {
m_distribution = newDstr.getSelectedTag().getID();
}
}
/**
* Returns the current distribution that'll be used for calculating the random
* matrix
*
* @return the current distribution
*/
public SelectedTag getDistribution() {
return new SelectedTag(m_distribution, TAGS_DSTRS_TYPE);
}
/**
* Returns the tip text for this property
*
* @return tip text for this property suitable for displaying in the
* explorer/experimenter gui
*/
public String replaceMissingValuesTipText() {
return "If set the filter uses weka.filters.unsupervised.attribute.ReplaceMissingValues"
+ " to replace the missing values instead of just skipping them.";
}
/**
* Sets either to use replace missing values filter or not
*
* @param t if true then the replace missing values is used
*/
public void setReplaceMissingValues(boolean t) {
m_useReplaceMissing = t;
}
/**
* Gets the current setting for using ReplaceMissingValues filter
*
* @return true if the replace missing values filter is used
*/
public boolean getReplaceMissingValues() {
return m_useReplaceMissing;
}
/**
* Returns the Capabilities of this filter.
*
* @return the capabilities of this object
* @see Capabilities
*/
@Override
public Capabilities getCapabilities() {
Capabilities result = super.getCapabilities();
result.disableAll();
// attributes
result.enableAllAttributes();
result.enable(Capability.MISSING_VALUES);
result.disable(Capability.STRING_ATTRIBUTES);
result.disable(Capability.RELATIONAL_ATTRIBUTES);
// class
result.enable(Capability.NUMERIC_CLASS);
result.enable(Capability.DATE_CLASS);
result.enable(Capability.NOMINAL_CLASS);
result.enable(Capability.MISSING_CLASS_VALUES);
result.enable(Capability.NO_CLASS);
return result;
}
/**
* Sets the format of the input instances.
*
* @param instanceInfo an Instances object containing the input instance
* structure (any instances contained in the object are ignored -
* only the structure is required).
* @return true if the outputFormat may be collected immediately
* @throws Exception if the input format can't be set successfully
*/
@Override
public boolean setInputFormat(Instances instanceInfo) throws Exception {
super.setInputFormat(instanceInfo);
if (instanceInfo.classIndex() >= 0) {
m_ntob = new weka.filters.supervised.attribute.NominalToBinary();
} else {
m_ntob = new weka.filters.unsupervised.attribute.NominalToBinary();
}
m_replaceMissing = null;
m_OutputFormatDefined = false;
if (getReplaceMissingValues()) {
m_replaceMissing = new ReplaceMissingValues();
m_replaceMissing.setInputFormat(instanceInfo);
if (m_ntob.setInputFormat(m_replaceMissing.getOutputFormat())) {
setOutputFormat();
return true;
} else {
return false;
}
}
if (m_ntob.setInputFormat(instanceInfo)) {
setOutputFormat();
return true;
} else {
return false;
}
}
/**
* Input an instance for filtering.
*
* @param instance the input instance
* @return true if the filtered instance may now be collected with output().
* @throws IllegalStateException if no input format has been set
*/
@Override
public boolean input(Instance instance) throws Exception {
if (getInputFormat() == null) {
throw new IllegalStateException("No input instance format defined");
}
if (m_NewBatch) {
resetQueue();
m_NewBatch = false;
}
if ((m_OutputFormatDefined) && ((!m_useReplaceMissing) || isFirstBatchDone())) {
if (m_replaceMissing != null) {
m_replaceMissing.input(instance);
instance = m_replaceMissing.output();
}
m_ntob.input(instance);
instance = m_ntob.output();
push(convertInstance(instance), false);
return true;
}
bufferInput(instance);
return false;
}
/**
* Signify that this batch of input to the filter is finished.
*
* @return true if there are instances pending output
* @throws NullPointerException if no input structure has been defined,
* @throws Exception if there was a problem finishing the batch.
*/
@Override
public boolean batchFinished() throws Exception {
if (getInputFormat() == null) {
throw new NullPointerException("No input instance format defined");
}
Instances insts = getInputFormat();
if (m_useReplaceMissing) {
insts = Filter.useFilter(insts, m_replaceMissing);
}
insts = Filter.useFilter(insts, m_ntob);
if (!m_OutputFormatDefined) {
setOutputFormat();
}
for (Instance instance : insts) {
push(convertInstance(instance), false); // No need to copy
}
flushInput();
m_NewBatch = true;
m_FirstBatchDone = true;
return (numPendingOutput() != 0);
}
/** Sets the output format */
protected void setOutputFormat() throws Exception {
Instances currentFormat = m_ntob.getOutputFormat();
if (m_percent > 0) {
m_k = (int) ((getInputFormat().numAttributes() - 1) * m_percent);
}
ArrayList attributes = new ArrayList();
for (int i = 0; i < m_k; i++) {
attributes.add(new Attribute("K" + (i + 1)));
}
if (getInputFormat().classIndex() > -1) {
attributes.add((Attribute) getInputFormat().classAttribute().copy());
}
Instances newFormat = new Instances(currentFormat.relationName(), attributes, 0);
if (getInputFormat().classIndex() > -1) {
newFormat.setClassIndex(attributes.size() - 1);
}
m_random = new Random(m_rndmSeed);
m_rmatrix = new double[m_k][currentFormat.numAttributes()];
if (m_distribution == GAUSSIAN) {
for (int i = 0; i < m_rmatrix.length; i++) {
for (int j = 0; j < m_rmatrix[i].length; j++) {
m_rmatrix[i][j] = m_random.nextGaussian();
}
}
} else {
boolean useDstrWithZero = (m_distribution == SPARSE1);
for (int i = 0; i < m_rmatrix.length; i++) {
for (int j = 0; j < m_rmatrix[i].length; j++) {
m_rmatrix[i][j] = rndmNum(useDstrWithZero);
}
}
}
m_OutputFormatDefined = true;
setOutputFormat(newFormat);
}
/**
* converts a single instance to the required format
*
* @param instance the instance to convert
* @return the converted instance
*/
protected Instance convertInstance(Instance instance) throws Exception {
double vals[] = new double[outputFormatPeek().numAttributes()];
for (int j = 0; j < m_k; j++) {
for (int i = 0; i < instance.numValues(); i++) {
int index = instance.index(i);
if (index != instance.classIndex()) {
double value = instance.valueSparse(i);
if (!Utils.isMissingValue(value)) {
vals[j] += m_rmatrix[j][index] * value;
}
} else {
vals[m_k] = instance.valueSparse(i);
}
}
}
return new DenseInstance(instance.weight(), vals);
}
private static final int weights[] = { 1, 1, 4 };
private static final int vals[] = { -1, 1, 0 };
private static final int weights2[] = { 1, 1 };
private static final int vals2[] = { -1, 1 };
private static final double sqrt3 = Math.sqrt(3);
/**
* returns a double x such that
* x = sqrt(3) * { -1 with prob. 1/6, 0 with prob. 2/3, 1 with prob. 1/6 }
*
* @param useDstrWithZero
* @return the generated number
*/
protected double rndmNum(boolean useDstrWithZero) {
if (useDstrWithZero) {
return sqrt3 * vals[weightedDistribution(weights)];
} else {
return vals2[weightedDistribution(weights2)];
}
}
/**
* Calculates a weighted distribution
*
* @param weights the weights to use
* @return
*/
protected int weightedDistribution(int[] weights) {
int sum = 0;
for (int weight : weights) {
sum += weight;
}
int val = (int) Math.floor(m_random.nextDouble() * sum);
for (int i = 0; i < weights.length; i++) {
val -= weights[i];
if (val < 0) {
return i;
}
}
return -1;
}
/**
* Returns the revision string.
*
* @return the revision
*/
@Override
public String getRevision() {
return RevisionUtils.extract("$Revision: 14609 $");
}
/**
* Main method for testing this class.
*
* @param argv should contain arguments to the filter: use -h for help
*/
public static void main(String[] argv) {
runFilter(new RandomProjection(), argv);
}
}
Ads help maintain this website.