This preview shows page 1-2-3-4-5-6-7-51-52-53-54-55-56-57-58-103-104-105-106-107-108-109 out of 109 pages.
Data Stream Classification and Novel Class DetectionOutline of The PresentationIntroductionData Stream ClassificationData Stream Classification (cont..)ChallengesInfinite LengthConcept-DriftConcept-EvolutionDynamic FeaturesSlide 11Slide 12DataStream Classification (cont..)OverviewEnsemble of ClassifiersEnsemble Classification of Data StreamsConcept-Evolution ProblemECSMiner: OverviewAlgorithmNovel Class DetectionTraining: Creating Decision BoundaryOutlier Detection and FilteringSlide 23Computing Cohesion & SeparationSpeeding UpAlgorithm To Detect Novel Class“Speedup” PenaltySlide 28Experiments - DatasetsExperiments - SetupExperiments - BaselineExperiments - ResultsSlide 33Slide 34Experiments – Parameter SensitivityExperiments – RuntimeSlide 37Feature Mapping Across Models and Test Data PointsFeature Space Conversion – Lossy-L Mapping (Local)Feature Space Conversion – Lossy-L MappingConversion Strategy II – Lossy-L MappingConversion Strategy III – D-Preserving MappingSlide 43Slide 44DiscussionComparisonSlide 47Slide 48Slide 49Baseline ApproachesApproaches ComparisonExperiments: DatasetsExperiments: ResultsSlide 54Experiments: SetupExperiments: BaselineTwitter ResultsSlide 58NASA DatasetForest Cover ResultsSlide 61KDD ResultsSlide 63Summary ResultsImproved Outlier Detection and Multiple Novel Class DetectionOutlier Threshold (OUTTH)Slide 69Slide 70Slide 71Dynamic threshold settingSlide 73Defer approach and Eager approach comparisonOutliers StatisticsSlide 76Outlier Statistics Gini AnalysisSlide 78Outlier Statistics Gini Analysis LimitationSlide 80Multi Novel Class DetectionSlide 82Slide 83Slide 84Slide 85Slide 86Slide 87Slide 88Slide 89Slide 90Slide 91Slide 92Slide 93Slide 94Slide 95Slide 96Slide 97Slide 98Result SummarySlide 100Running Time ComparisonMulti Novel Detection ResultsSlide 103ConclusionReferencesReferences (contd.)Slide 107Slide 108Slide 109Slide 110Slide 111University of Texas at DallasData Stream Classification and Data Stream Classification and Novel Class DetectionNovel Class DetectionMehedy Masud, Latifur Khan, Qing Chen and Bhavani ThuraisinghamDepartment of Computer Science , University of Texas at DallasJing Gao, Jiawei HanDepartment of Computer Science , University of Illionois at Urbana ChampaignCharu AggarwalIBM T. J. WatsonThis work was funded in part by Aug 10, 2011Masud et al.University of Texas at DallasOutline of The PresentationOutline of The PresentationBackgroundData Stream ClassificationNovel Class DetectionAug 10, 2011Masud et al. 2University of Texas at DallasIntroductionIntroductionCharacteristics of Data streams are: ◦Continuous flow of dataNetwork trafficSensor dataCall center records◦Examples:Aug 10, 2011Masud et al. 3University of Texas at DallasUses past labeled data to build classification modelPredicts the labels of future instances using the modelHelps decision making Data Stream ClassificationData Stream ClassificationNetwork trafficClassification modelAttack trafficFirewallBlock and quarantineBenign trafficServerModel updateExpert analysis and labelingAug 10, 2011Masud et al. 4University of Texas at DallasData Stream Classification Data Stream Classification (cont..)(cont..)What are the applications?◦Security Monitoring◦Network monitoring and traffic engineering.◦Business : credit card transaction flows.◦Telecommunication calling records.◦Web logs and web page click streams.Aug 10, 2011Masud et al. 5University of Texas at DallasInfinite lengthConcept-driftConcept-evolutionFeature EvolutionChallengesChallengesAug 10, 2011Masud et al. 6University of Texas at DallasImpractical to store and use all historical data◦Requires infinite storage ◦And running timeInfinite LengthInfinite LengthAug 10, 2011Masud et al. 7University of Texas at DallasConcept-DriftConcept-DriftNegative instancePositive instanceA data chunkCurrent hyperplanePrevious hyperplaneInstances victim of concept-driftAug 10, 2011Masud et al. 8University of Texas at DallasConcept-EvolutionConcept-Evolution X X X X X X X X X X XX X X X X X XX X X X X X X X X X X X X X X X XX X X X X XX X X X X X Novel classyx1y1y2x ++++ ++ ++ + + ++ + +++ ++ + ++ + + + ++ + +++++ ++++ +++ + ++ + + ++ ++ + - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - -- - - - - Classification rules: R1. if (x > x1 and y < y2) or (x < x1 and y < y1) then class = +R2. if (x > x1 and y > y2) or (x < x1 and y > y1) then class = -Existing classification models misclassify novel class instancesACDByx1y1y2x ++++ ++ ++ + + ++ + +++ ++ + ++ + + + ++ + +++++ ++++ +++ + ++ + + ++ ++ + - - - - - - - - - - - - - - - + + + + + + + + + + + + + + + + - - - - - - - - - - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - -- - - - - - - - - - - - - - - - - - - - - - -- - - - - ACDBAug 10, 2011Masud et al. 9University of Texas at DallasDynamic FeaturesDynamic FeaturesWhy new features evolving◦Infinite data streamNormally, global feature set is unknownNew features may appear◦Concept driftAs concept drifting, new features may appear◦Concept evolutionNew type of class normally holds new set of featuresDifferent chunks may have different feature setsAug 10, 2011Masud et al. 10University of Texas at DallasDynamic FeaturesDynamic FeaturesFeature Extraction & Selectioni + 1st chunkith chunk Existing classification models need complete fixed features and apply to all the chunks. Global features are difficult to predict. One solution is using all English words and generate vector. Dimension of the vector will be too high.Current modelTraining New ModelFeature SpaceConversionClassification &Novel Class Detectionrunway, climbrunway, clear, ramprunway, ground, rampith chunk and i + 1st chunk and models have different feature setsFeature SetAug 10, 2011Masud et al. 11University of Texas at DallasOutline of The PresentationOutline of The PresentationIntroductionData Stream ClassificationNovel Class DetectionAug 10, 2011Masud et al. 12University of Texas at DallasDataStream Classification DataStream Classification (cont..) (cont..) Single Model Incremental
View Full Document
Unlocking...