moa.classifiers.meta

Class OzaBagASHT

java.lang.Object

moa.AbstractMOAObject

moa.options.AbstractOptionHandler

moa.classifiers.AbstractClassifier

moa.classifiers.meta.OzaBag

moa.classifiers.meta.OzaBagASHT

All Implemented Interfaces:

Serializable, Classifier, AWTRenderable, MOAObject, OptionHandler

public class OzaBagASHT
extends OzaBag

Bagging using trees of different size. The Adaptive-Size Hoeffding Tree (ASHT) is derived from the Hoeffding Tree algorithm with the following differences:

• it has a maximum number of split nodes, or size
• after one node splits, if the number of split nodes of the ASHT tree is higher than the maximum value, then it deletes some nodes to reduce its size

The intuition behind this method is as follows: smaller trees adapt more quickly to changes, and larger trees do better during periods with no or little change, simply because they were built on more data. Trees limited to size s will be reset about twice as often as trees with a size limit of 2s. This creates a set of different reset-speeds for an ensemble of such trees, and therefore a subset of trees that are a good approximation for the current rate of change. It is important to note that resets will happen all the time, even for stationary datasets, but this behaviour should not have a negative impact on the ensemble’s predictive performance. When the tree size exceeds the maximun size value, there are two different delete options:

• delete the oldest node, the root, and all of its children except the one where the split has been made. After that, the root of the child not deleted becomes the new root
• delete all the nodes of the tree, i.e., restart from a new root.

The maximum allowed size for the n-th ASHT tree is twice the maximum allowed size for the (n-1)-th tree. Moreover, each tree has a weight proportional to the inverse of the square of its error, and it monitors its error with an exponential weighted moving average (EWMA) with alpha = .01. The size of the first tree is 2.

With this new method, it is attempted to improve bagging performance by increasing tree diversity. It has been observed that boosting tends to produce a more diverse set of classifiers than bagging, and this has been cited as a factor in increased performance.
See more details in:

Albert Bifet, Geoff Holmes, Bernhard Pfahringer, Richard Kirkby, and Ricard Gavaldà. New ensemble methods for evolving data streams. In 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2009.

The learner must be ASHoeffdingTree, a Hoeffding Tree with a maximum size value.

Example:

OzaBagASHT -l ASHoeffdingTree -s 10 -u -r Parameters:

• Same parameters as OzaBag
• -f : the size of first classifier in the bag.
• -u : Enable weight classifiers
• -e : Reset trees when size is higher than the max

Version:

$Revision: 7 $

Author:

Albert Bifet (abifet at cs dot waikato dot ac dot nz)

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
protected double	alpha
protected double[]	error
IntOption	firstClassifierSizeOption
FlagOption	resetTreesOption
FlagOption	useWeightOption

Fields inherited from class moa.classifiers.meta.OzaBag

baseLearnerOption, ensemble, ensembleSizeOption

Fields inherited from class moa.classifiers.AbstractClassifier

classifierRandom, modelContext, randomSeed, randomSeedOption, trainingWeightSeenByModel

Fields inherited from class moa.options.AbstractOptionHandler

classOptionNamesToPreparedObjects, options

Constructor Summary

Constructors

Constructor and Description
OzaBagASHT()

Method Summary

Methods

Modifier and Type	Method and Description
void	getModelDescription(StringBuilder out, int indent) Returns a string representation of the model.
String	getPurposeString() Gets the purpose of this object
double[]	getVotesForInstance(weka.core.Instance inst) Predicts the class memberships for a given instance.
void	resetLearningImpl() Resets this classifier.
void	trainOnInstanceImpl(weka.core.Instance inst) Trains this classifier incrementally using the given instance. The reason for ...Impl methods: ease programmer burden by not requiring them to remember calls to super in overridden methods.

Methods inherited from class moa.classifiers.meta.OzaBag

getModelMeasurementsImpl, getSubClassifiers, isRandomizable

Methods inherited from class moa.classifiers.AbstractClassifier

contextIsCompatible, copy, correctlyClassifies, getAttributeNameString, getAWTRenderer, getClassLabelString, getClassNameString, getDescription, getModelContext, getModelMeasurements, getNominalValueString, modelAttIndexToInstanceAttIndex, modelAttIndexToInstanceAttIndex, prepareForUseImpl, resetLearning, setModelContext, setRandomSeed, trainingHasStarted, trainingWeightSeenByModel, trainOnInstance

Methods inherited from class moa.options.AbstractOptionHandler

discoverOptionsViaReflection, getCLICreationString, getOptions, getPreparedClassOption, getPreparedClassOption, prepareClassOptions, prepareForUse, prepareForUse

Methods inherited from class moa.AbstractMOAObject

copy, measureByteSize, measureByteSize, toString

Methods inherited from class java.lang.Object

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

Methods inherited from interface moa.options.OptionHandler

getCLICreationString, getOptions, prepareForUse, prepareForUse

Methods inherited from interface moa.MOAObject

measureByteSize

Field Detail

firstClassifierSizeOption

public IntOption firstClassifierSizeOption

useWeightOption

public FlagOption useWeightOption

resetTreesOption

public FlagOption resetTreesOption

error

protected double[] error

alpha

protected double alpha

Constructor Detail

OzaBagASHT

public OzaBagASHT()

Method Detail

getPurposeString

public String getPurposeString()

Description copied from interface: OptionHandler

Gets the purpose of this object

Specified by:

getPurposeString in interface OptionHandler

Overrides:

getPurposeString in class OzaBag

Returns:

the string with the purpose of this object

resetLearningImpl

public void resetLearningImpl()

Description copied from class: AbstractClassifier

Resets this classifier. It must be similar to starting a new classifier from scratch.

The reason for ...Impl methods: ease programmer burden by not requiring them to remember calls to super in overridden methods. Note that this will produce compiler errors if not overridden.

Overrides:

resetLearningImpl in class OzaBag

trainOnInstanceImpl

public void trainOnInstanceImpl(weka.core.Instance inst)

Description copied from class: AbstractClassifier

Trains this classifier incrementally using the given instance.

The reason for ...Impl methods: ease programmer burden by not requiring them to remember calls to super in overridden methods. Note that this will produce compiler errors if not overridden.

Overrides:

trainOnInstanceImpl in class OzaBag

Parameters:

inst - the instance to be used for training

getVotesForInstance

public double[] getVotesForInstance(weka.core.Instance inst)

Description copied from interface: Classifier

Predicts the class memberships for a given instance. If an instance is unclassified, the returned array elements must be all zero.

Specified by:

getVotesForInstance in interface Classifier

Overrides:

getVotesForInstance in class OzaBag

Parameters:

inst - the instance to be classified

Returns:

an array containing the estimated membership probabilities of the test instance in each class

getModelDescription

public void getModelDescription(StringBuilder out,
                       int indent)

Description copied from class: AbstractClassifier

Returns a string representation of the model.

Overrides:

getModelDescription in class OzaBag

Parameters:

out - the stringbuilder to add the description

indent - the number of characters to indent