com.yahoo.egads.models.adm

Class AdaptiveKernelDensityChangePointDetector

java.lang.Object

com.yahoo.egads.models.adm.AnomalyDetectionAbstractModel

com.yahoo.egads.models.adm.AdaptiveKernelDensityChangePointDetector

All Implemented Interfaces:

JsonAble, Model, AnomalyDetectionModel, java.io.Serializable

public class AdaptiveKernelDensityChangePointDetector
extends AnomalyDetectionAbstractModel

AdaptiveKernelDensityChangePointDetector implements density-based algorithm for change point detection. Input: 1. the residual time-series = actual - expected 2. PRE_WINDOW_SIZE: the size of PRE_WINDOW 3. POST_WINDOW_SIZE: the size of POST_WINDOW 4. CONFIDENCE: the confidence level at which the threshold for the KL-Divergence score is computed. Output: the time indices when the distribution of residual changes significantly In particular, the algorithm slides two side-by-side windows (PRE_WINDOW and POST_WINDOW) and computes the KL-divergence between the distribution of the residuals in two windows for each time index. The distribution of the residuals in each window is computed non-parametrically using Kernel Density Estimation with Gaussian kernels with adaptive bandwidths. Once the KL-divergence is computed for each time index, it is thresholded (where the threshold is directly proportional to 'CONFIDENCE'). And finally for each continuous segment of the thresholded KL-divergence time-series, the index of the maximum value is reported as a change point.

See Also:

Serialized Form

Constructor Summary

Constructors

Constructor and Description
AdaptiveKernelDensityChangePointDetector(java.util.Properties config)

Method Summary

Modifier and Type	Method and Description
Anomaly.IntervalSequence	detect(TimeSeries.DataSequence observedSeries, TimeSeries.DataSequence expectedSeries)
java.util.ArrayList<java.lang.Integer>	detectChangePoints(float[] residuals, int preWindowSize, int postWindowSize, float confidence)
void	fromJson(JSONObject json_obj)
java.lang.String	getModelName()
java.lang.String	getType()
void	reset()
void	toJson(JSONStringer json_out)
void	tune(TimeSeries.DataSequence observedSeries, TimeSeries.DataSequence expectedSeries, Anomaly.IntervalSequence anomalySequence)

Methods inherited from class com.yahoo.egads.models.adm.AnomalyDetectionAbstractModel

getModelType, isDetectionWindowPoint

Methods inherited from class java.lang.Object

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

Constructor Detail

AdaptiveKernelDensityChangePointDetector

public AdaptiveKernelDensityChangePointDetector(java.util.Properties config)

Method Detail

toJson

public void toJson(JSONStringer json_out)
            throws java.lang.Exception

Specified by:

toJson in interface JsonAble

Overrides:

toJson in class AnomalyDetectionAbstractModel

Throws:

java.lang.Exception

fromJson

public void fromJson(JSONObject json_obj)
              throws java.lang.Exception

Specified by:

fromJson in interface JsonAble

Overrides:

fromJson in class AnomalyDetectionAbstractModel

Throws:

java.lang.Exception

getType

public java.lang.String getType()

getModelName

public java.lang.String getModelName()

Specified by:

getModelName in interface Model

Overrides:

getModelName in class AnomalyDetectionAbstractModel

reset

public void reset()

tune

public void tune(TimeSeries.DataSequence observedSeries,
                 TimeSeries.DataSequence expectedSeries,
                 Anomaly.IntervalSequence anomalySequence)
          throws java.lang.Exception

Throws:

java.lang.Exception

detect

public Anomaly.IntervalSequence detect(TimeSeries.DataSequence observedSeries,
                                       TimeSeries.DataSequence expectedSeries)
                                throws java.lang.Exception

Throws:

java.lang.Exception

detectChangePoints

public java.util.ArrayList<java.lang.Integer> detectChangePoints(float[] residuals,
                                                                 int preWindowSize,
                                                                 int postWindowSize,
                                                                 float confidence)