Cone penetrationtestdatahasbeenwidelyusedfordeterminationofthethresholdofseismically inducedsoilliquefaction.However,possibleinaccuraciesinthecollecteddatafromcasehistoriesas well asnaturalvariabilityofparametersandotheruncertaintiesassociatedwithnaturalphenomenon have yetprohibitedaconclusivedefinitionforthisthreshold. Variousclassificationtechniqueshavebeenusedtodefinethemostreliablecorrelations.However, availableliquefiedtonon-liquefieddataimbalancehascausedlearningbiastothemajorityclassinthe learning modelofthepatternrecognitionsystems.Thishasadverselyaffectedtheoutcomeofsuch approachesandinordertoovercomethisproblemSupportVectorDataDescription(SVDD)strategyis employedto‘‘upsample’’theminoritydata.InotherwordsSVDD,whichisrobustagainstnoisy samples,isusedtogeneratevirtualdatapointsfortheminorityclass,bearingthesamecharacteristics as thenon-virtualsamples.Inordertospecifythemostappropriatedatarangeasphereboundary around themainbodyofthedataaresoughtthroughanoptimizationprocess.Thedatainsidethe obtainedboundaryarethetargetdataandtheonesoutsideitaretheoutliersorso-called‘‘noise’’,tobe neglected.Thisprocedurereducestheissueofclassintermixtureinthefringezoneandproduces relativelywelldefinedclassthatthenisfedintotheAdaptiveNeuro-FuzzyInferenceSystem(ANFIS) classifierfordeterminationofliquefactionpotential.Thepredictionsarethenexaminedtoevaluatethe reliabilityandvalidationoftheoveralltechniqueandcomparedwithotherpredictionmethodsusing confusionmatrix.Itisshownthattheoverallaccuracyoftheproposedtechniqueishigherthanall previouslyproposedmethodsandonlyequaltotheSupportVectorMachine(SVM)technique. FurthermoreanimprovementintheF-scoreofthenon-liquefieddatarecognitionhasbeenachieved in relationtoallpreviouslyproposedmethods.

Keywords