基于决策树和神经网络的交通事故分析(IJITCS-V6-N2-3)

I.J.InformationTechnologyandComputerScience,2014,02,22-28PublishedOnlineJanuary2014inMECS()DOI:10.5815/ijitcs.2014.02.03Copyright©2014MECSI.J.InformationTechnologyandComputerScience,2014,02,22-28TrafficAccidentAnalysisUsingDecisionTreesandNeuralNetworksOlutayoV.AComputerScienceDepartment,JosephAyoBabalolaUniversity,Ikeji-Arakeji,NigeriaCorrespondingE-mail:[e-mail]nt,JosephAyoBabalolaUniversity,Ikeji-Arakeji,NigeriaE-mail:[e-mail]—ThisworkemployedArtificialNeuralNetworksandDecisionTreesdataanalysistechniquestodiscovernewknowledgefromhistoricaldataaboutaccidentsinoneofNigeria’sbusiestroadsinordertoreducecarnageonourhighways.Dataofaccidentsrecordsonthefirst40kilometresfromIbadantoLagoswerecollectedfromNigeriaRoadSafetyCorps.Thedatawereorganizedintocontinuousandcategoricaldata.ThecontinuousdatawereanalysedusingArtificialNeuralNetworkstechniqueandthecategoricaldatawerealsoanalysedusingDecisionTreestechnique.Sensitivityanalysiswasperformedandirrelevantinputswereeliminated.TheperformancemeasuresusedtodeterminetheperformanceofthetechniquesincludeMeanAbsoluteError(MAE),ConfusionMatrix,AccuracyRate,TruePositive,FalsePositiveandPercentagecorrectlyclassifiedinstances.Experimentalresultsrevealthat,betweenthemachineslearningparadigmsconsidered,DecisionTreeapproachoutperformedtheArtificialNeuralNetworkwithalowererrorrateandhigheraccuracyrate.Ourresearchanalysisalsoshowsthat,thethreemostimportantcausesofaccidentareTyreburst,lossofcontrolandoverspeeding.IndexTerms—TrafficAccidentDataMining,AccidentCausesPredictionandSensitivityAnalysis,PerformanceComparisonI.IntroductionTheproblemofdeathsandinjuriesasaresultofaccidentsisacknowledgedtobeaglobalphenomenonandtrafficsafetyhasbeenaseriousconcernsincethestartoftheautomobileage,almostonehundredyearsago.Ithasbeenestimatedthatover300,000personsdieand10to15millionpersonsareinjuredeveryyearinroadaccidentsthroughouttheworld.Statisticshavealsoshownthatmortalityinroadaccidentsisveryhighamongyoungadultsthatconstitutethemajorpartoftheworkforce.Inordertocombatthisproblem,variousroadsafetystrategies,methodsandcountermeasureshavebeenproposedandused.Thesemethodsmainlyinvolveconsciousplanning,designandoperationsofroads.Oneimportantfeatureofthismethodistheidentificationandtreatmentofaccident-pronelocationscommonlycalledblackspots[1].However,blackspotsarenottheonlycausesofaccidentsonthehighway.Regressionanalysisisacommonapproachusedinmodellinghighwaygeometrics,trafficcharacterizationsandaccidentfrequenciesinordertodetermineothercausesofaccidents.RegressionanalysishighlydependsontrafficflowdatasuchasAverageDailyTraffic(ADT).Italsorequirestheresearchertoknowexactlythedependentvariablesaswellastheindependentvariables.Sadlyhowever,inNigeria,dataareoftenlookedatfromonedimension.Moreoftenthannot,thecausesforroadaccidentsindevelopingcountrylikeNigeriamayhavenothingtodowiththehighwaygeometry,oreventrafficcharacterization.Also,alargenumberofdataminingalgorithmicsolutionexist;butuntilnow,littleornoempiricalresearchhasbeendoneoncomparingtheirefficiencyespeciallyonroadaccidentsdataset.Therefore,thisresearchworkisusefultoascertainwhichofthesedataminingclassification’salgorithmicsolutionswillscalebetter(ArtificialNeuralNetworksandDecisionTrees)onroadaccidentdatabase.Finally,thepurposeofthisresearchistolookathistoricaldataofroadaccidentsononeoftheNigeria’sbusiestroadsonhowcanbemoreanalysedinordertodiscovernewknowledgeaboutroadaccidentsinNigeriaandusethisknowledgetoreducethecarnageonourhighway.Relatedimportantworkscanbesummarizedasfollows.Abdelwahabetal.[2]studiedthe1997accidentdatafortheCentralFloridaareafocusingontwo-vehicleaccidentsthatoccurredatsignalizedintersections.Theinjuryseveritywasdividedintothreeclasses:noinjury,possibleinjuryanddisablinginjury.TheperformanceofNeuralNetwork(NN)trainedbyLevenberg-MarquardtalgorithmandFuzzyARTMAPwerecompared,andfoundthatNN(65.6%and60.4%TrafficAccidentAnalysisUsingDecisionTreesandNeuralNetworks23Copyright©2014MECSI.J.InformationTechnologyandComputerScience,2014,02,22-28classificationaccuracyforthetrainingandtestingphases)performedbetterthanFuzzyARTMAP(56.1%).Bedardetal.[3]appliedamultivariatelogisticregressiontodeterminetheindependentcontributionofdriver,crash,andvehiclecharacteristicstodrivers’fatalityrisk.Itwasfoundthatincreasingseatbeltuse,reducingspeed,andreducingthenumberandseverityofdriver-sideimpactsmightpreventfatalities.Someresearchersstudiedtherelationshipbetweendrivers’age,gender,vehiclemass,impactspeedordrivingspeedmeasurewithfatalities[4,]Diaetal.usedreal-worlddatafordevelopingamulti-layeredNNfreewayincidentdetectionmodel[5].ResultsshowedthatNNcouldprovidefasterandmorereliableincidentdetectionoverthemodelthatwasinoperationonMelbourne’sfreeways.Evancoconductedamultivariatepopulation-basedstatisticalanalysistodeterminetherelationshipbetweenfatalitiesandaccidentnotificationtimes[6].Theanalysisdemonstratedthataccidentnotificationtimeisanimportantdeterminantofthenumberoffata