Comment
Author: Admin | 2025-04-28
Fraudulentsource.The goal of detectionsoftware is to classify eachtransaction as legal orfraudulent. The types of errors that can occur in this classification areclassification of a fraudulent transaction as legal (false negative); andclassificationofalegaltransactionasfraudulent(falsepositive).Probabilityofdetection=PD=Pr(classifyintoH1│H1istrue)orProbabilityoffalsenegative=1–PDProbabilityoffalsepositive=PF=Pr(classifyintoH1│H0istrue)LetthenumericalvaluesforthenormalandfraudulenttransactionsfollowexponentialdistributionswithparametersλNandλF,λN>λFrespectively.TheprobabilityofdetectionPDandprobabilityoffalsepositivePFasttxFDFFedxePλλλ−∞−==∫ttxNFNNedxePλλλ−∞−==∫ThusPDcanbeexpressedasafunctionofPFasrFDPP =wherer=λF/λNisbetween0and1.Trees [18] stated that the quality profile of most detection software ischaracterizedbyacurvethatrelatesitsPDandPF known asthe receiver operatingcharacteristic curve (ROC). ROC is a function that summarizes the possibleperformancesofadetector.Itvisualizesthetrade-offbetweenfalsealarmratesanddetectionrates,thusfacilitatingthechoiceofadecisionfunctions.ThiswasthereforeusedintheperformanceanalysisofCCFandotherdetectionwatchsystems.The effectiveness of this detection software is measured in terms of theclassification errors, which consist of systemdetectionrate and false alarm rate.The data used in the application were collected from a Nigerian bank, whichconsistoftransactiondata madeperday during theobserved period, thatis,inamonth. The collection was done according to the two types of bank operationsinvestigated. The aim is to identify frauds within each of the categories byidentifyingflaggedtransactions usingtheCCF detection watchapproach.DetailsofthetestdatasetsarelistedinTable1.Table1.SummaryoftheTwoDataSubsetsUsedtoTesteachoftheOperationsofCCFDetectionModel.Operation Transaction Fraudulent ProportionoffraudulentWithdrawal 10,650 5 0.47%Deposit 8,102 2 0.24%Total 18,752 7 0.37% DataMiningApplicationinCreditCardFraudDetectionSystem317JournalofEngineeringScienceandTechnologyJune2011,Vol.6(3)The performance analyses of the respective detection algorithms are car riedoutusingMATLABsoftwarepackageandtheresultscomparedwiththecollecteddataareasshowninFigs.4and5.Fig.4.ROCforWithdrawalFraudDetection.Fig.5.ROCforDepositFraudDetection.Figures4and5showthatthemodelresults comparedsatisfactorilywellwiththecollecteddataresultsforthetwotransactionsexaminedinthispaper.Inthe comparisonof CCF detection watch performance with twootherfrauddetection models using the ROC curve, it wa sno ted that the results from CCFdetection using neural network was accurate and reliable. The two differentcommercial products, quadratic discriminates analysis (QDA) and logisticregression (LOGIT) wereselected to test the feasibilityofusing neural networktools for the purpose of CCF detection watch. The performance anal ysis of theCCF detection watch model (deposit transaction) is compared with these twocommercialpackagesandtheresultisshowninFig.5.FromFig.6,theROCcurveforCCFdetectionwatchdetectsover95%offraudcaseswithoutcausingfalsealarms.ThisisfollowedbytheROCcurveforlogisticregression with 75% detection with no false alarms, the quadratic discriminant 318F.N.OgwuelekaJournalofEngineeringScienceandTechnologyJune2011,Vol.6(3)analysis with 60% detection. This shows that the performance of CCF detectionwatchisinagreementwithotherdetectionsoftware,butperformsbetter.Fig.6.ComparisonofCCFDetectionWatchwithotherFraudDetectionSystemROCforDepositFraudDetection.Using two software products enabled this work to illustrate different userinterfaces available, alternative to neural networks mod els, design for decision-making, and the performance metrics of different models. Comparison ofperformanceofneuralnetworkmodelwithtraditionalstatisticalmodelsincreasedthe confidence in the ability of CCFdetection watc hin successful modelling ofcreditcardfraudsinthebankingindustry.In the comparison of the performance evaluation of the two- stage clustermodelandthe four-stagecluster model,thetwo-stagecredit card frauddetectionmodel works as a binary classifier that has two choices, “fraud” or “legal”. Itworks by producing a score. The score is a measure of the confidence of theclassifierthataparticular transactionislegal or fraud. To decide whichitis,thescoreiscomparedtoathresholdthenitisdeemedtobelegal.Ifitisgreater,thenit is fraud. This type of model consists of a component and five user interfacesandtends to classify most legaltransactions as fraudulent. Thefour-stage creditcardfrauddetection modeldesignedconsistsof twocomponentswithseven userinterfaces.ThetwocomponentsareArtificialNeuralNetwork(ANN)componentand the Rule Based (RB) component. It was developed using four classes ofcluster-low,high,riskyandhighrisk.IntheSOMNNengine,thedatasetdescription,numberofdatapointsin thedataset,andthenumberofclustersareentered.Thedatapoint’sentrywascreatedandfilled.Theentriesaresorted inascending order under the cluster list. Whenthetrain/generateclustersareselected,the clusterlabelbecomesreadyforfillingdependingonthenumberofclustersenteredandthenstoredonthedatabase.Thedatabase was stored in Microsoft Access table and was also used to determinewhena card transactionwastobe processed,blocked,unblocked,orthe alertsetoff.After each transaction, the datapoint entry and clusters madeareprocessedbytheSOMNNengineandsentintothedatabase.Thishelpsthedetectionenginetoknowwhenanydataentryislegitimateorfraudulent,andthe reasonis givenimmediately after the alert. The transaction will not be processed, but will be DataMiningApplicationinCreditCardFraudDetectionSystem319JournalofEngineeringScienceandTechnologyJune2011,Vol.6(3)storedinthedatabase.TheSOMNNsetupreportdisplaysthenumberofclusters,thenamesoftheclustersandthelistofclustertransactions.The database was meant to run at the background of the existing bankingsoftware and be getting its data from real-time banking transaction, checkingwhether the transaction is legitimate (and so will be processed) or fraudulent(transaction will not be processed and the alert will be let off with reasonsdisplayed). The detection software i si nterfaced with the source data. MicrosoftAccess was used to create and manage the database. Each of the data tablescreatedinthisstudysuchastransactiontables,accountopentables,clustertables,legitimate and fraudulent tables, is stored in a separate file. In the cluster nametable, four cluster names were used. Since the
Add Comment