com.ibm.bluej.consistency.classifier

Class AbstractClassifier

java.lang.Object

com.ibm.bluej.consistency.classifier.AbstractClassifier

All Implemented Interfaces:

java.io.Serializable

Direct Known Subclasses:

LibLinearClassifier, LibSVMClassifier

public abstract class AbstractClassifier
extends java.lang.Object
implements java.io.Serializable

See Also:

Serialized Form

Field Summary

Fields

Modifier and Type	Field and Description
static com.ibm.bluej.util.common.MutableDouble	PRESENT

Constructor Summary

Constructors

Constructor and Description
AbstractClassifier()

Method Summary

Methods

Modifier and Type	Method and Description
static <T> java.util.Map<T,com.ibm.bluej.util.common.MutableDouble>	convertBinaryFeatures(java.lang.Iterable<T> features) Converts a set of binary features into the feature/value map format expected by train
<T> com.ibm.bluej.util.common.PrecisionRecallThreshold	crossValidate(java.util.List<java.util.Map<T,com.ibm.bluej.util.common.MutableDouble>>[] allVectors) Ten fold cross validation
<T> com.ibm.bluej.util.common.PrecisionRecallThreshold	crossValidate(java.util.List<java.util.Map<T,com.ibm.bluej.util.common.MutableDouble>>[] allVectors, int folds, long seed) If binary classifier it assumes the "relevant" class for P/R is index 0
static <S,M extends java.util.Map<S,com.ibm.bluej.util.common.MutableDouble>> java.util.Collection<M>[]	holdOut(java.util.Collection<M>[] allVectors, com.ibm.bluej.util.common.IdentitySet<M> heldOut)
abstract int	predict(java.util.Map<?,com.ibm.bluej.util.common.MutableDouble> features) Predict the most likely class label
abstract double[]	predictProb(java.util.Map<?,com.ibm.bluej.util.common.MutableDouble> features) Predict a probability for each class
static double	probToWeight(double prob)
static java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[]	pruneFeatures(java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[] trainingData, int minOccurrences)
static java.util.ArrayList<java.util.Map<java.lang.Integer,com.ibm.bluej.util.common.MutableDouble>>[]	testingData(int numClasses, int numInstances, int numFeatures, int numNoise) Generate some data for testing, if numFeatures and numInstances are pretty high it should be possible to do well on this data
abstract void	train(java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[] allVectors) Trains the model using the allVectors as training data

Methods inherited from class java.lang.Object

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

Field Detail

PRESENT

public static final com.ibm.bluej.util.common.MutableDouble PRESENT

Constructor Detail

AbstractClassifier

public AbstractClassifier()

Method Detail

convertBinaryFeatures

public static <T> java.util.Map<T,com.ibm.bluej.util.common.MutableDouble> convertBinaryFeatures(java.lang.Iterable<T> features)

Converts a set of binary features into the feature/value map format expected by train

Parameters:

features - binary features

Returns:

map of the feature to a MutableDouble with value 1.0

pruneFeatures

public static java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[] pruneFeatures(java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[] trainingData,
                                                                                                       int minOccurrences)

testingData

public static java.util.ArrayList<java.util.Map<java.lang.Integer,com.ibm.bluej.util.common.MutableDouble>>[] testingData(int numClasses,
                                                                                                          int numInstances,
                                                                                                          int numFeatures,
                                                                                                          int numNoise)

Generate some data for testing, if numFeatures and numInstances are pretty high it should be possible to do well on this data

Parameters:

numClasses -

numInstances -

numFeatures -

numNoise -

Returns:

probToWeight

public static double probToWeight(double prob)

holdOut

public static <S,M extends java.util.Map<S,com.ibm.bluej.util.common.MutableDouble>> java.util.Collection<M>[] holdOut(java.util.Collection<M>[] allVectors,
                                                                                                       com.ibm.bluej.util.common.IdentitySet<M> heldOut)

crossValidate

public <T> com.ibm.bluej.util.common.PrecisionRecallThreshold crossValidate(java.util.List<java.util.Map<T,com.ibm.bluej.util.common.MutableDouble>>[] allVectors,
                                                                   int folds,
                                                                   long seed)

If binary classifier it assumes the "relevant" class for P/R is index 0

Parameters:

allVectors -

folds -

crossValidate

public <T> com.ibm.bluej.util.common.PrecisionRecallThreshold crossValidate(java.util.List<java.util.Map<T,com.ibm.bluej.util.common.MutableDouble>>[] allVectors)

Ten fold cross validation

Parameters:

allVectors -

train

public abstract void train(java.util.Collection<? extends java.util.Map<?,com.ibm.bluej.util.common.MutableDouble>>[] allVectors)

Trains the model using the allVectors as training data

Parameters:

allVectors - The indices of the array correspond to the classes to be distinguished. Each class has an Iterable of instances - feature/value maps.

predictProb

public abstract double[] predictProb(java.util.Map<?,com.ibm.bluej.util.common.MutableDouble> features)

Predict a probability for each class

Parameters:

features - The features for the instance to be classified

Returns:

The probability for each class

predict

public abstract int predict(java.util.Map<?,com.ibm.bluej.util.common.MutableDouble> features)

Predict the most likely class label

Parameters:

features - The features for the instance to be classified

Returns:

The class label