ai.libs.jaicore.ml.tsc.classifier.trees

Class TimeSeriesTreeLearningAlgorithm

java.lang.Object

ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,C>

ai.libs.jaicore.ml.tsc.classifier.ASimplifiedTSCLearningAlgorithm<java.lang.Integer,TimeSeriesTreeClassifier>

ai.libs.jaicore.ml.tsc.classifier.trees.TimeSeriesTreeLearningAlgorithm

All Implemented Interfaces:

ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>, ai.libs.jaicore.basic.Cancelable, ai.libs.jaicore.basic.ILoggingCustomizable, java.lang.Iterable<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>, java.util.concurrent.Callable<TimeSeriesTreeClassifier>, java.util.Iterator<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>

public class TimeSeriesTreeLearningAlgorithm
extends ASimplifiedTSCLearningAlgorithm<java.lang.Integer,TimeSeriesTreeClassifier>

Algorithm to build a time series tree as described in Deng, Houtao et al. "A Time Series Forest for Classification and Feature Extraction." Inf. Sci. 239 (2013): 142-153.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static interface	TimeSeriesTreeLearningAlgorithm.ITimeSeriesTreeConfig

Field Summary

Fields

Modifier and Type	Field and Description
static double	ENTROPY_APLHA Alpha parameter used to weight the importance of the feature's margins to the threshold candidates.
static int	NUM_THRESH_CANDIDATES Number of threshold candidates created in each tree recursion step.
static boolean	USE_BIAS_CORRECTION Indicator that the bias (Bessel's) correction should be used for the calculation of the standard deviation.

Constructor Summary

Constructors

Constructor and Description
TimeSeriesTreeLearningAlgorithm(TimeSeriesTreeLearningAlgorithm.ITimeSeriesTreeConfig config, TimeSeriesTreeClassifier tree, TimeSeriesDataset data) Constructor.

Method Summary

Modifier and Type	Method and Description
static double	calculateDeltaEntropy(double[] dataValues, int[] targets, double thresholdCandidate, java.util.List<java.lang.Integer> classes, double parentEntropy) Function calculating the delta entropy for a given thresholdCandidate and parentEntropy.
static double	calculateEntrance(double deltaEntropy, double margin) Calculates the entrance gain specified by Deng et. al. in the paper's chapter 4.1.
static double	calculateMargin(double[] dataValues, double thresholdCandidate) Function calculating the margin between the given thresholdCandidate and the nearest feature value from the given dataValues.
TimeSeriesTreeClassifier	call() Training procedure construction a time series tree using the given input data.
void	cancel()
static java.util.List<java.util.List<java.lang.Double>>	generateThresholdCandidates(ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> pairOfIntervalLists, int numOfCandidates, double[][][] transformedFeatures) Function generating threshold candidates for each feature type.
int	getBestSplitIndex(double[] deltaEntropyStarPerFeatureType) Function returning feature type used for the split based on given the deltaEntropy star values.
static ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>>	getChildDataIndices(double[][][] transformedData, int n, int fType, int t1t2, double threshold) Function returning the data indices assigned to the left and the right child of a binary tree based on the splitting criterion given by the feature type fType, the intervals index t1t2 in the transformed data set transformedData and the threshold.
int	getNumCPUs()
ai.libs.jaicore.basic.TimeOut	getTimeout()
boolean	hasNext()
java.util.Iterator<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>	iterator()
ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent	next()
ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent	nextWithException()
static java.util.List<java.lang.Integer>	randomlySampleNoReplacement(java.util.List<java.lang.Integer> list, int sampleSize, int seed) Function sampling a given list randomly without replacement using the given seed.
void	registerListener(java.lang.Object listener)
static ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>>	sampleIntervals(int m, int seed) Function sampling intervals based on the length of the time series m and the given seed.
void	setNumCPUs(int numberOfCPUs)
void	setTimeout(long timeout, java.util.concurrent.TimeUnit timeUnit)
void	setTimeout(ai.libs.jaicore.basic.TimeOut timeout)
double[][][]	transformInstances(double[][] dataset, ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> pairOfItervalLists) Method transforming the given dataset using the interval pairs specified in T1T2 by calculating each TimeSeriesFeature.FeatureType for every instance and interval pair.
void	tree(double[][] data, int[] targets, double parentEntropy, ai.libs.jaicore.graph.TreeNode<ai.libs.jaicore.ml.tsc.classifier.trees.TimeSeriesTreeClassifier.TimeSeriesTreeNodeDecisionFunction> nodeToBeFilled, int depth) Tree generation (cf.

Methods inherited from class ai.libs.jaicore.ml.tsc.classifier.ASimplifiedTSCLearningAlgorithm

getClassifier

Methods inherited from class ai.libs.jaicore.basic.algorithm.AAlgorithm

activate, announceTimeoutDetected, avoidReinterruptionOnShutdownOnCurrentThread, checkAndConductTermination, checkTermination, computeTimeoutAware, getActivationTime, getConfig, getId, getInput, getLoggerName, getRemainingTimeToDeadline, getState, getTimeoutPrecautionOffset, hasThreadBeenInterruptedDuringShutdown, interruptThreadAsPartOfShutdown, isCanceled, isShutdownInitialized, isStopCriterionSatisfied, isTimeouted, post, registerActiveThread, resolveShutdownInterruptOnCurrentThread, setConfig, setLoggerName, setMaxNumThreads, setState, setTimeoutPrecautionOffset, shutdown, terminate, unregisterActiveThread, unregisterThreadAndShutdown

Methods inherited from class java.lang.Object

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

Methods inherited from interface java.lang.Iterable

forEach, spliterator

Methods inherited from interface java.util.Iterator

forEachRemaining, remove

Field Detail

NUM_THRESH_CANDIDATES

public static final int NUM_THRESH_CANDIDATES

Number of threshold candidates created in each tree recursion step.

See Also:

Constant Field Values

ENTROPY_APLHA

public static final double ENTROPY_APLHA

Alpha parameter used to weight the importance of the feature's margins to the threshold candidates.

See Also:

Constant Field Values

USE_BIAS_CORRECTION

public static final boolean USE_BIAS_CORRECTION

Indicator that the bias (Bessel's) correction should be used for the calculation of the standard deviation.

See Also:

Constant Field Values

Constructor Detail

TimeSeriesTreeLearningAlgorithm

public TimeSeriesTreeLearningAlgorithm(TimeSeriesTreeLearningAlgorithm.ITimeSeriesTreeConfig config,
                                       TimeSeriesTreeClassifier tree,
                                       TimeSeriesDataset data)

Constructor.

Parameters:

maxDepth - Maximal depth of the tree to be trained

seed - Seed used for randomized operations

useFeatureCaching - Indicator whether feature caching should be used. Since feature generation is very efficient, this should be only used if the time series is very long

Method Detail

registerListener

public void registerListener(java.lang.Object listener)

Specified by:

registerListener in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

registerListener in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

getNumCPUs

public int getNumCPUs()

Specified by:

getNumCPUs in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

getNumCPUs in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

setNumCPUs

public void setNumCPUs(int numberOfCPUs)

Specified by:

setNumCPUs in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

setNumCPUs in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

setTimeout

public void setTimeout(long timeout,
                       java.util.concurrent.TimeUnit timeUnit)

Specified by:

setTimeout in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

setTimeout in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

setTimeout

public void setTimeout(ai.libs.jaicore.basic.TimeOut timeout)

Specified by:

setTimeout in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

setTimeout in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

getTimeout

public ai.libs.jaicore.basic.TimeOut getTimeout()

Specified by:

getTimeout in interface ai.libs.jaicore.basic.algorithm.IAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

Overrides:

getTimeout in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

nextWithException

public ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent nextWithException()

call

public TimeSeriesTreeClassifier call()

Training procedure construction a time series tree using the given input data.

iterator

public java.util.Iterator<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent> iterator()

Specified by:

iterator in interface java.lang.Iterable<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>

Overrides:

iterator in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

hasNext

public boolean hasNext()

Specified by:

hasNext in interface java.util.Iterator<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>

Overrides:

hasNext in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

next

public ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent next()

Specified by:

next in interface java.util.Iterator<ai.libs.jaicore.basic.algorithm.events.AlgorithmEvent>

Overrides:

next in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

cancel

public void cancel()

Specified by:

cancel in interface ai.libs.jaicore.basic.Cancelable

Overrides:

cancel in class ai.libs.jaicore.basic.algorithm.AAlgorithm<TimeSeriesDataset,TimeSeriesTreeClassifier>

tree

public void tree(double[][] data,
                 int[] targets,
                 double parentEntropy,
                 ai.libs.jaicore.graph.TreeNode<ai.libs.jaicore.ml.tsc.classifier.trees.TimeSeriesTreeClassifier.TimeSeriesTreeNodeDecisionFunction> nodeToBeFilled,
                 int depth)

Tree generation (cf. Algorithm 2 of original paper). Samples the intervals in each recursion step and calculates the features (using caches if TimeSeriesTreeLearningAlgorithm#useFeatureCaching was set true). It then searches for an optimal split regarding several threshold candidates for feature splits. The splitting criterion is based on a metric called Entrance gain which is a combination of the entropy induced by the class proportions and the feature margins to the threshold (cf. chapter 4.1 in the paper). The tree's recursion is stopped at a leaf node if there is no entropy gain, the maxDepth has been reached or the local entropy is zero.

Parameters:

data - The untransformed data which will be used for the split in the transformed feature representation

targets - The targets of the instances

parentEntropy - The parent entropy calculated in the recursion's previous step

nodeToBeFilled - The tree node which should be filled with the splitting information to use it for predictions

depth - The current depth to be compared to the TimeSeriesTreeLearningAlgorithm#maxDepth

getChildDataIndices

public static ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> getChildDataIndices(double[][][] transformedData,
                                                                                                                                       int n,
                                                                                                                                       int fType,
                                                                                                                                       int t1t2,
                                                                                                                                       double threshold)

Function returning the data indices assigned to the left and the right child of a binary tree based on the splitting criterion given by the feature type fType, the intervals index t1t2 in the transformed data set transformedData and the threshold.

Parameters:

transformedData - Transformed data on which the split is calculated

n - The number of instances

fType - The feature type to be used for the split

t1t2 - The interval's index in the transformedData to be used for the split

threshold - The threshold to be used for the split

Returns:

Returns a pair of two lists, storing the data indices for the data points assigned to the left child of the current node (X) and the data indices assigned to the right child (Y)

getBestSplitIndex

public int getBestSplitIndex(double[] deltaEntropyStarPerFeatureType)

Function returning feature type used for the split based on given the deltaEntropy star values. If multiple feature types have generated the same deltaEntropy value, a random decision is taken.

Parameters:

deltaEntropyStarPerFeatureType - The delta entropy star value per feature

Returns:

Returns the feature type index which has been chosen

calculateDeltaEntropy

public static double calculateDeltaEntropy(double[] dataValues,
                                           int[] targets,
                                           double thresholdCandidate,
                                           java.util.List<java.lang.Integer> classes,
                                           double parentEntropy)

Function calculating the delta entropy for a given thresholdCandidate and parentEntropy. The values of the data are the feature type's values for each instance. The delta entropy is formed of the difference between the parent entropy and the weighted sum of the entropy values of the children and their instance assignments based on the split.

Parameters:

dataValues - The transformed feature type values for each instance

targets - The targets of each instance

thresholdCandidate - The threshold candidate to be evaluated

classes - List storing the classes whose indices can be looked up

parentEntropy - The parent entropy used for the delta calculation

Returns:

Returns the delta entropy for the threshold candidate of the current feature type

calculateEntrance

public static double calculateEntrance(double deltaEntropy,
                                       double margin)

Calculates the entrance gain specified by Deng et. al. in the paper's chapter 4.1.

Parameters:

deltaEntropy - The delta entropy

margin - The features margin

Returns:

Returns the entrance gain

calculateMargin

public static double calculateMargin(double[] dataValues,
                                     double thresholdCandidate)

Function calculating the margin between the given thresholdCandidate and the nearest feature value from the given dataValues.

Parameters:

dataValues - The feature values compared to the candidate

thresholdCandidate - The threshold candidate which is assessed

Returns:

Returns the minimum distance among the feature values and the threshold candidate

transformInstances

public double[][][] transformInstances(double[][] dataset,
                                       ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> pairOfItervalLists)

Method transforming the given dataset using the interval pairs specified in T1T2 by calculating each TimeSeriesFeature.FeatureType for every instance and interval pair.

Parameters:

dataset - The dataset which should be transformed

pairOfItervalLists - The start and end interval pairs (see sampleIntervals(int, int))

Returns:

Returns the transformed instances (shape: number of feature types x number of interval pairs x number of instances)

generateThresholdCandidates

public static java.util.List<java.util.List<java.lang.Double>> generateThresholdCandidates(ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> pairOfIntervalLists,
                                                                                           int numOfCandidates,
                                                                                           double[][][] transformedFeatures)

Function generating threshold candidates for each feature type. It calculates the interval [min f_k(t1,t2), max f_k(t1,t2)] among all instances for every feature type and every possible interval and generates numberOfCandidates candidates using equal-width intervals.

Parameters:

pairOfIntervalLists - The pair of start and end interval pairs (see sampleIntervals(int, int))

numOfCandidates - The number of candidates to be generated per feature type

transformedFeatures - The transformed data instances

Returns:

Returns a list consisting of a list for each feature type storing the threshold candidates

sampleIntervals

public static ai.libs.jaicore.basic.sets.Pair<java.util.List<java.lang.Integer>,java.util.List<java.lang.Integer>> sampleIntervals(int m,
                                                                                                                                   int seed)

Function sampling intervals based on the length of the time series m and the given seed. Refers to algorithm 1 of the paper. The sampled intervals are stored in a pair of lists where each index of the first list is related to the same index in the second list. Sampling is done without replacement.

Parameters:

m - Number of time series attributes (steps)

seed - The seed used for the randomized sampling

Returns:

Returns a pair of lists consisting of the start indices (X) and the end indices (Y)

randomlySampleNoReplacement

public static java.util.List<java.lang.Integer> randomlySampleNoReplacement(java.util.List<java.lang.Integer> list,
                                                                            int sampleSize,
                                                                            int seed)

Function sampling a given list randomly without replacement using the given seed. sampleSize many elements are sampled and returned.

Parameters:

list - List to be sampled from without replacement

sampleSize - Number of elements to be sampled (must be <= list.size())

seed - The seed used for the randomized sampling

Returns:

Returns a list of elements which have been sampled