ASK-the-Expert: Active learning based knowledge discovery ...€¦ · ASK-the-Expert: Active...

Post on 09-Oct-2020

3 views 0 download

Transcript of ASK-the-Expert: Active learning based knowledge discovery ...€¦ · ASK-the-Expert: Active...

ASK-the-Expert:Activelearningbasedknowledgediscoveryusingthe expert

Kamalika DasDataSciences Group

NASAAmesResearch Center

MLWorkshop,August 2017

Problem• Identifysafetyeventsinflightoperational data• Unsupervisedanomaly detection• SMEreviewof anomalies

Unsupervisedanomalydetection

NOS OSOS NOS

NOS NOS

Statisticalflight anomalies2

• Lackofdefinitionof‘safety’ incident• One-classSVMbasedanomaly detection

x2 x2

Θ

Unsupervisedanomaly detection

x1 x1

S. Das, B. Matthews, A. Srivastava, N Oza. 2010.Multiple kernel learning for heterogeneous anomaly detection: algorithm and aviation safety case study. InProceedings of the 16th ACM SIGKDD (KDD '10). 47-56.

3

Stateofthe art

TRACON

ATRCC

FAAFacilities

Data Collection Data Processing

TRACON

ARTCC

DataFilter

FeatureSelectionAndNormalization

DataMerge

Calculate FlightSeparation andTurn-to-finalfeatures

Existing System

MKAD: UnsupervisedAnomaly Detection Nominals

Anomalies

Labels

OperationallySignificantEvents

4

Proposed approach

Input Features Anomalies

MKAD

Operationally significant anomalies

Active learning strategy

SMETraining

Uninteresting anomalies

Nominals

Active learning with rationales framework

Inst

ance

forl

abel

ing

Labe

l

5

Ratio

nale

Output

2-class classification/ranking algorithmActive Learner

Activelearning framework

Flightsflight1

f?1 f2 f3 … … … … fn f*?

?

?

labels

Statistical anomaliesFeatures

Bootstrap samples

Labeled poolf1 f2 f3 … … … … fn f*

O

N

N

f1 f2 f3 … … … … fn f*?

?

?

?

Unlabeled pool

ActiveLearner

Model:2-classmultiplekernelSVMActivelearningstrategy:MostlikelypositiveAutomatedfeatureconstruction:Multiplekernellearning+decisiontree construction

Lossof separationOSflight:x* Label:y* rationale

flight: x* Sample to label

6

ASK-the-Experttool: architecture

7K.Das,I.Avrekh,B.Matthews,M.Sharma,N.Oza.2017.ASK-the-Expert:Activelearningbasedknowledgediscoveryusingthe expert.InProceedingsofECML-PKDD2017.Tobepublished.

Annotator component

8

Coordinator component

10

Multiplekernelsupportvector machine

• 2-classSVM objective:

• Decision function:

f1 f2 f2 … … fn

1

2

3

m

• Multiplekernel2classSVM:classifyingbetweenoperationallysignificant(OS)anduninteresting(NOS) flights

Feature set

Flight

timeserie

s

… Weightedaverageofallfeature kernels

… …

ηnKernelweights: η1 … η3 ……

10

Rationalefeatureconstruction

• Howtosetweights:η1,η2,…,ηn𝑠. 𝑡. 𝜂𝑚 >=0 &∑𝜂𝑚= 1

• SimpleMKL algorithm– Modifiedobjective function– Alternatesbetweenoptimizingclassifiermarginandweightsof kernels

11

Rationalefeatureconstruction

• Decisiontree induction

12

Data

altitude

Verticalseparation

Horizontal separation

ORIGINAL FEATURES• Latitude• Longitude• Altitude• Ground speed• Horizontal separation• Vertical separation• Aircraft size• Turn-to-final(TTF) parameters:

• Maximum overshoot• Speedat TTF• Distanceat TTF• Angleat TTF• Altitudedifferenceat TTF

• Nearestneighboring(NN)flight info:• NNflightonsame runway• NNflightonparallel runway• NNflightpartofthesame flow

Runway

Rationale features

“Lossof separation”• Horizontalseparation<3milesAND

Verticalseparation<1000ftANDnearestneighboringflightisnotonparallelrunwaysandnotpartofthesame flow

“Large overshoot”• Maximumovershootisgreaterthana

thresholdbasedonvaluesofflightswithpositive labels

“Unusualflight path”• Overalldeviationfromexpected(average)

trajectoryofalllandingflightsonthatrunway

x

Begin PointxLandingPoint

Expected

Actutrajecto

alry

trajectory

Deviation fromexpectedpath

Verticalseparation<1000 ft

Horizontalseparation<3miles

Experimental setup

15

• Dataset:30NMairspacearoundDenverInternationalAirportforAug 2014– Trainingset:~2400 flights– Statisticalanomalies: 153– OSflights: 24

• 2foldcrossvalidationwith10randombootstrapsforeach fold

Performance analysis• Metrics:precision@5and precision@10• Most-likelypositivestrategy

Learningcurvesfordifferentactivelearning strategies16

Performance analysis

75%savingsinlabelingeffort

Learningcurvesformostlikelypositivestrategywithandwithout rationales

17M.Sharma,K.Das,M.Bilgic,B.Matthews,D.Nielsen,N.Oza.2016.ActiveLearningwithRationalesforIdentifyingOperationallySignificantAnomaliesinAviation.InProceedingsofECML-PKDD2016.pp209-225.

Performance analysis

Comparisonofnumberoflabeledflightsrequiredbyvariousstrategiestoachieveatargetperformancemeasure.‘n/a’representsthatthetarget

performancecannotbeachievedbyamethodevenwith45labeled flights.

18

Performance benefits

20

• Generalization– Twodifferenttestdatasets:July2014andJuly 2015– Averageimprovementinprecision@5: ~30%– Averageimprovementinprecision@10: ~65%

• Review time– Upto75%reductioninreviewtimeforsametargetperformance

Summary

20

• Upto75%reductioninSMEreview time• Methodandtoolisagnostictodomain

• Canbetailoredtoworkinanydomainsufferingfromlackoflabeleddata

Acknowledgement

21

• ThisworkissupportedbyCenterInnovationFund(CIF)2017 award

• Team:– NikunjOza,NASAAmesResearch Center– BryanMatthews,SGT Inc.– IllyaAvrekh,SGT Inc.– ManaliSharma,PhDStudent,IllinoisInstituteof Technology– SayeriLala,UndergraduateStudent,MassachusettsInstituteof Technology

Thank You

22