大数据数据挖掘培训讲义_偏差检测

《大数据数据挖掘培训讲义_偏差检测》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、SummarizationandDeviationDetection--Whatisnew?2OutlineSummarizationKEFIR–KeyFindingsReporterWSARE–WhatisStrangeAboutRecentEvents3WhatisNew?Olddatanewdata4SummarizationConciselysummarizewhatisnewanddifferent,unexpectedwithrespecttopreviousvalueswithrespecttoexpectedvalues…Focusonwhatisa

2、ctionable!5Problem:HealthcareCostsHealthcarecostsinUS:1outof7GDP$andrisingpotentialproblems:fraud,misuse,…understandingwheretheproblemsareisfirststeptofixingthemGTE–selfinsuredformedicalcostsGTEhealthcarecosts–$X00,000,000Task:Analyzeemployeehealthcaredataandgenerateareportthatdescribest

3、hemajorproblems6GTEKeyFindingsReporter:KEFIRKEFIRApproach:AnalyzeallpossibledeviationsSelectinterestingfindingsAugmentkeyfindingswith:ExplanationsofplausiblecausesRecommendationsofappropriateactionsConvertfindingstoauser-friendlyreportwithtextandgraphicsKEFIRSearchSpace8Drill-DownExampl

4、e9WhatChangeIsImportant?10DeviationDetectionDrillDownthroughthesearchspaceGenerateafindingforeachmeasuredeviationfrompreviousperioddeviationfromnormdeviationprojectedfornextperiod,ifnoactionInterestingnessofDeviationsImpact:howmuchthedeviationaffectsthebottomlineSavingsPercentage:howmuch

5、ofthedeviationfromthenormcanbeexpectedtobesavedbytheactionRecommendationsHierarchicalrecommendationrulesdefineappropriateinterventionstrategiesforimportantmeasuresandstudyareas.Example:measure=admissionrateper1000&study_area=Inpatientadmissions&percent_change>0.10IfThenUtilizationreviewi

6、sneededintheareaofadmissioncertification.ExpectedSavings:20%13ExplanationAmeasureisexplainedbyfindingthepathofrelatedmeasureswiththehighestimpactThelargeincreaseinm1ingroups1wascausedbyanincreaseinm3,whichwascausedbyariseinm5,primarilyinsectors13.14ReportGenerationAutomaticgenerationofbu

7、siness-user-orientedreportsNaturallanguagegenerationwithtemplatematchingGraphicsdeliveredviabrowser16SampleKEFIRpagesOverviewInpatientadmissionsStatusPrototypeimplementedinGTEin1995KEFIRreceivedGTE’shighestawardfortechnicalachievementin1995KeybusinessuserleftGTEin1996ands