欢迎来到天天文库
浏览记录
ID:52421020
大小:5.33 MB
页数:44页
时间:2020-04-06
《数据挖掘之异常检测.ppt》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、AnomalyDetection:AintroductionSourceofslides:TutorialAtAmericanStatisticalAssociation(ASA2008)JiaweiHan-datamining:conceptsandtechniquesTutorialattheEuropeanConferenceonPrinciplesandPracticeofKnowledgeDiscoveryinDatabasesSpeaker:WentaoLiOutlineDefinitionApplicationMethodsLim
2、itedtime,SoIjustdrawthepictureofanomalydetection,formoredetail,pleaseturntothepaperforhelp.WhatareAnomalies?AnomalyisapatterninthedatathatdoesnotconformtotheexpectedbehaviorAnomalyisAdataobjectthatdeviatessignificantlyfromthenormalobjectsasifitweregeneratedbyadifferentmechan
3、ismAlsoreferredtoasoutliers,exceptions,peculiarities,surprises,etc.Anomaliestranslatetosignificant(oftencritical)reallifeentitiesCyberintrusionsCreditcardfraudFaultsinmechanicalsystemsRelatedproblemsOutliersaredifferentfromthenoisedataNoiseisrandomerrororvarianceinameasuredv
4、ariableNoiseshouldberemovedbeforeoutlierdetectionOutliersareinteresting:ItviolatesthemechanismthatgeneratesthenormaldataOutlierdetectionvs.noveltydetection:earlystage,outlier;butlatermergedintothemodelKeyChallengesDefiningarepresentativenormalregionischallengingTheboundarybe
5、tweennormalandoutlyingbehaviorisoftennotpreciseAvailabilityoflabeleddatafortraining/validationTheexactnotionofanoutlierisdifferentfordifferentapplicationdomainsDatamightcontainnoiseNormalbehaviorkeepsevolvingAppropriateselectionofrelevantfeaturesMapRelatedareas(theory)Applic
6、ation(practice)ProblemformulationDetectioneffect+AspectsofAnomalyDetectionProblemNatureofinputdataWhatisthecharacteristicofinputdataAvailabilityofsupervisionNumberoflabelTypeofanomaly:point,contextual,structuralTypeofanomalyOutputofanomalydetectionScorevslabelEvaluationofano
7、malydetectiontechniquesWhatkindofdetectionisgoodInputDataMostcommonformofdatahandledbyanomalydetectiontechniquesisRecordDataUnivariateMultivariateInputDataMostcommonformofdatahandledbyanomalydetectiontechniquesisRecordDataUnivariateMultivariateInputData–NatureofAttributesNat
8、ureofattributesBinaryCategoricalContinuousHybridcategoricalcontinuouscontin
此文档下载收益归作者所有