featRep
Class FeatureRepresentation

java.lang.Object
  featRep.FeatureRepresentation

public class FeatureRepresentation
extends java.lang.Object

This class converts records in relational format to a feature vector based on various heuristics described in the following paper:

 @inproceedings{DBLP:conf/icde/InanKB09,
 author    = {Ali Inan and
             Murat Kantarcioglu and
             Elisa Bertino},
 title     = {Using Anonymized Data for Classification},
 booktitle = {ICDE},
 year      = {2009},
 pages     = {429-440}
 }
 

The class attribute is assumed to be the last attribute of the input file. Please replace if that is not the case.

The constructor parameters are as follows:

• A descriptor file that specifies attribute types (and domains for categorical atts)
• An anonymization configuration file
• Numeric feature representation heuristic (see Params)
• Categorical feature representation heuristic (see Params)
• If expected distance will be used, the method of computation (i.e., either assuming all values are distributed uniformly over the domain or using QI-statistics)

Field Summary

java.lang.Object[]	attributes Vector of attribute (either NumericAtt or CategoricalAtt object
private int	catRep Categorical representation heuristic of choice.
private java.lang.String[]	classValues Vector of class values
private int[]	featAttMapping The i^th entry contains the index of the attribute feature i correspond to
private java.util.LinkedList<java.lang.Integer>	idAttributes List of identifier attribute indices within the original file
boolean[]	isCont Entry i is true if i^th attribute is numerical
private boolean[]	isGeneralized Entry i is true if i^th attribute has been generalized
private int	numRep Numeric representation heuristic of choice.
private QIDAttribute[]	qids Quasi-identifier attributes (indices should be adjusted according to id attributes)
private boolean	usePDFExp Choice of expected distance calculation method.
private boolean	useUniExp Choice of expected distance calculation method.

Constructor Summary

FeatureRepresentation(java.lang.String descriptorFilename, QIDAttribute[] qids, java.util.LinkedList<java.lang.Integer> idAttributes, int numRep, int catRep, boolean usePDFExp, boolean useUniExp)
Class constructor

Method Summary

void	assignFeatureIndices() Assigns feature indices to each attribute
double	dotProduct(svm_node[] x) Computes the expected dot product
double	dotProduct(svm_node[] x, svm_node[] y) Computes the expected dot product
private double	dotProductPDF(svm_node[] x) Computes the expected dot product based on QI-statistics
private double	dotProductPDF(svm_node[] x, svm_node[] y) Computes the expected dot product based on QI-statistics
private double	dotProductUni(svm_node[] x) Computes the expected dot product based on the assumption that values of a generalization are distributed uniformly
private double	dotProductUni(svm_node[] x, svm_node[] y) Computes the expected dot product based on the assumption that values of a generalization are distributed uniformly
void	featurize(java.lang.String inputFile, java.lang.String outputFile) Convert the input to a set of feature vectors, which will be written to the output file
void	initialScan(java.lang.String inputFile) Sets lower/upper bound for numeric attributes (does nothing for categorical)
private boolean	isIdentifier(int index) Checks if the attribute at index is an identifier
private void	readDescriptor(java.lang.String descriptor) Read the descriptor file
double	squareDistance(svm_node[] x, svm_node[] y) Computes the expected square distance
private double	squareDistancePDF(svm_node[] x, svm_node[] y) Computes the expected square distance based on QI-statistics
private double	squareDistanceUni(svm_node[] x, svm_node[] y) Computes the expected square distance based on the assumption that values of a generalization are distributed uniformly

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

classValues

private java.lang.String[] classValues

Vector of class values

attributes

public java.lang.Object[] attributes

Vector of attribute (either NumericAtt or CategoricalAtt object

isCont

public boolean[] isCont

Entry i is true if i^th attribute is numerical

isGeneralized

private boolean[] isGeneralized

Entry i is true if i^th attribute has been generalized

numRep

private int numRep

Numeric representation heuristic of choice. See Params for alternatives.

catRep

private int catRep

Categorical representation heuristic of choice. See Params for alternatives.

usePDFExp

private boolean usePDFExp

Choice of expected distance calculation method. This one uses the PDF obtained from an anonymization outputted with the genValsDist output format option of the anonymization toolbox.

useUniExp

private boolean useUniExp

Choice of expected distance calculation method. This one assumes all values within a generalization are uniformly distributed.

qids

private QIDAttribute[] qids

Quasi-identifier attributes (indices should be adjusted according to id attributes)

idAttributes

private java.util.LinkedList<java.lang.Integer> idAttributes

List of identifier attribute indices within the original file

featAttMapping

private int[] featAttMapping

The i^th entry contains the index of the attribute feature i correspond to

Constructor Detail

FeatureRepresentation

public FeatureRepresentation(java.lang.String descriptorFilename,
                             QIDAttribute[] qids,
                             java.util.LinkedList<java.lang.Integer> idAttributes,
                             int numRep,
                             int catRep,
                             boolean usePDFExp,
                             boolean useUniExp)

Class constructor

Parameters:

descriptorFilename - A names files that describes attribute types and/or domains

anonConfig - Anonymization configuration file

numRep - Numeric feature representation heuristic

catRep - Categorical feature representation heuristic

usePDFExp - Flag for calculating expected distance based on QI-statistics

useUniExp - Flag for calculating expected distance based on uniform distribution assumption

Method Detail

isIdentifier

private boolean isIdentifier(int index)

Checks if the attribute at index is an identifier

Parameters:

index - attribute index

Returns:

true if identifier, false otherwise

readDescriptor

private void readDescriptor(java.lang.String descriptor)

Read the descriptor file

Parameters:

descriptor - filename

initialScan

public void initialScan(java.lang.String inputFile)

Sets lower/upper bound for numeric attributes (does nothing for categorical)

Parameters:

inputFile - Input data filename

assignFeatureIndices

public void assignFeatureIndices()

Assigns feature indices to each attribute

featurize

public void featurize(java.lang.String inputFile,
                      java.lang.String outputFile)

Convert the input to a set of feature vectors, which will be written to the output file

Parameters:

inputFile - input filename

outputFile - output filename

squareDistance

public double squareDistance(svm_node[] x,
                             svm_node[] y)

Computes the expected square distance

Parameters:

x - vector of features

y - vector of features

Returns:

Expected square distance between x and y

dotProduct

public double dotProduct(svm_node[] x,
                         svm_node[] y)

Computes the expected dot product

Parameters:

x - vector of features

y - vector of features

Returns:

Expected dot product of x and y

dotProduct

public double dotProduct(svm_node[] x)

Computes the expected dot product

Parameters:

x - vector of features

Returns:

Expected dot product of x and x

squareDistancePDF

private double squareDistancePDF(svm_node[] x,
                                 svm_node[] y)

Computes the expected square distance based on QI-statistics

Parameters:

x - vector of features

y - vector of features

Returns:

Expected square distance between x and y

dotProductPDF

private double dotProductPDF(svm_node[] x,
                             svm_node[] y)

Computes the expected dot product based on QI-statistics

Parameters:

x - vector of features

y - vector of features

Returns:

Expected dot product between x and y

dotProductPDF

private double dotProductPDF(svm_node[] x)

Computes the expected dot product based on QI-statistics

Parameters:

x - vector of features

Returns:

Expected dot product between x and x

squareDistanceUni

private double squareDistanceUni(svm_node[] x,
                                 svm_node[] y)

Computes the expected square distance based on the assumption that values of a generalization are distributed uniformly

Parameters:

x - vector of features

y - vector of features

Returns:

Expected square distance between x and y

dotProductUni

private double dotProductUni(svm_node[] x,
                             svm_node[] y)

Computes the expected dot product based on the assumption that values of a generalization are distributed uniformly

Parameters:

x - vector of features

y - vector of features

Returns:

Expected dot product of x and y

dotProductUni

private double dotProductUni(svm_node[] x)

Computes the expected dot product based on the assumption that values of a generalization are distributed uniformly

Parameters:

x - vector of features

Returns:

Expected dot product of x and x