org.knime.base.node.mine.decisiontree2.learner2

Class SplitQualityGini

java.lang.Object

org.knime.base.node.mine.decisiontree2.learner2.SplitQualityMeasure

org.knime.base.node.mine.decisiontree2.learner2.SplitQualityGini

All Implemented Interfaces:

Cloneable

public class SplitQualityGini
extends SplitQualityMeasure

Implements the gini index split quality measure. This gini index is subtracted from 1 (worst value), thus the gini index is also better if it is larger than another gini index (same as for gain ratio).

Since:

2.6

Author:

Christoph Sieb, University of Konstanz

Constructor Summary

Constructors

Constructor and Description
SplitQualityGini()

Method Summary

Modifier and Type	Method and Description
double	getWorstValue() Returns the worst value for this quality measure.
void	initQualityMeasure(double[] classFrequencies, double allOverRecords) Some quality measures, like the information gain, calculate a quality of a previous distribution compared to a new one.
boolean	isBetter(double quality1, double quality2) A gini index is better if it is larger than the other one.
boolean	isBetterOrEqual(double quality1, double quality2) A GINI index is better if it is larger than the other one.
double	measureQuality(double allOverRecords, double[] partitionFrequency, double[][] partitionClassFrequency, double numUnknownRecords) Calculates the gini split index.
double	postProcessMeasure(double qualityMeasure, double allOverRecords, double[] partitionFrequency, double numUnknownRecords) The gini index need not to post process the measure.
String	toString()

Methods inherited from class org.knime.base.node.mine.decisiontree2.learner2.SplitQualityMeasure

clone

Methods inherited from class java.lang.Object

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

Constructor Detail

SplitQualityGini

public SplitQualityGini()

Method Detail

isBetter

public boolean isBetter(double quality1,
                        double quality2)

A gini index is better if it is larger than the other one. Determines if the first passed quality is better compared to the second quality.

Specified by:

isBetter in class SplitQualityMeasure

Parameters:

quality1 - first quality to compare

quality2 - second quality to compare

Returns:

true, iff the first quality is better to the second quality

isBetterOrEqual

public boolean isBetterOrEqual(double quality1,
                               double quality2)

A GINI index is better if it is larger than the other one. Determines if the first passed quality is better or equal compared to the second quality.

Specified by:

isBetterOrEqual in class SplitQualityMeasure

Parameters:

quality1 - first quality to compare

quality2 - second quality to compare

Returns:

true, iff the first quality is better or equal to the second quality

measureQuality

public double measureQuality(double allOverRecords,
                             double[] partitionFrequency,
                             double[][] partitionClassFrequency,
                             double numUnknownRecords)

Calculates the gini split index.

For a dataset T the gini index is: gini(T) = 1 - SUM(pj * pj) - for all relative class frequencies pj (pj = Pj/|T|). Pj is the absolut class frequency and nx the number of records in the data set

The gini for the split is: giniSplit(T) = SUM(nx/N*gini(Tx)) - for all relative partition frequencies nx/N and all partitions Tx

Specified by:

measureQuality in class SplitQualityMeasure

Parameters:

allOverRecords - the allover number of records with known values in the partition to split; corresponds to N in the formula

partitionFrequency - the frequencies of the different patitions; corresponds to nx in the formula

partitionClassFrequency - all class frequencies Pj (second dimension) for all partitions Tx (first dimension *

numUnknownRecords - the number of records with unknown (missing) value of the relevant attribute; used to weight the quality measure

Returns:

the gini split index

getWorstValue

public double getWorstValue()

Returns the worst value for this quality measure.

Specified by:

getWorstValue in class SplitQualityMeasure

Returns:

the worst value for this quality measure

initQualityMeasure

public void initQualityMeasure(double[] classFrequencies,
                               double allOverRecords)

Some quality measures, like the information gain, calculate a quality of a previous distribution compared to a new one. This previous distribution can be reused. For those cases a init method is provided that enable pre calculations to increase performance.

Specified by:

initQualityMeasure in class SplitQualityMeasure

Parameters:

classFrequencies - the class frequencies

allOverRecords - the overall count

toString

public String toString()

Specified by:

toString in class SplitQualityMeasure

postProcessMeasure

public double postProcessMeasure(double qualityMeasure,
                                 double allOverRecords,
                                 double[] partitionFrequency,
                                 double numUnknownRecords)

The gini index need not to post process the measure. Some quality measures need normalization when compared to other attributes. As this normalization is not required when the quality is compared inside a single attribute, this method allows to perform post processing (normalization) of quality measures to avoid a lot of unnecessary calculations.

Specified by:

postProcessMeasure in class SplitQualityMeasure

Parameters:

qualityMeasure - the quality measure to post process

allOverRecords - the allover number of known (non-missing) records

partitionFrequency - the frequencies of the potential split partitions

numUnknownRecords - the number of unknown (missing) records

Returns:

the post processed quality measure