混合聚类分类神经网络在反应性关节炎医学诊断中的应用(IJISA-V8-N8-1)

I.J.IntelligentSystemsandApplications,2016,8,1-9PublishedOnlineAugust2016inMECS()DOI:10.5815/ijisa.2016.08.01Copyright©2016MECSI.J.IntelligentSystemsandApplications,2016,8,1-9HybridClustering-ClassificationNeuralNetworkintheMedicalDiagnosticsoftheReactiveArthritisYevgeniyBodyanskiy,OlenaVynokurovaControlSystemsResearchLaboratory,KharkivNationalUniversityofRadioElectronics,Kharkiv,61166,UkraineE-mail:{yevgeniy.bodyanskiy,olena.vynokurova}@nure.uaVolodymyrSavvo,TatianaTverdokhlibPediatricsDepartment,KharkivMedicalAcademiaofPost-GraduateEducation,Kharkiv,61176,UkraineE-mail:[e-mail],[e-mail]sandAppliedMathematicsDepartment,UzhhorodNationalUniversity,Uzhhorod,88000,UkraineE-mail:[e-mail]—Inthepaper,thehybridclustering-classificationneuralnetworkisproposed.Thisnetworkallowstoincreaseaqualityofinformationprocessingundertheconditionofoverlappingclassesduetotherationalchoiceoflearningrateparameterandintroducingspecialprocedureoffuzzyreasoningintheclustering-classificationprocess,whichoccursbothwithexternallearningsignal(“supervised”),andwithoutone(“unsupervised”).Assimilaritymeasureneighborhoodfunctionormembershipone,cosinestructuresareused,whichallowtoprovideahighflexibilityduetoself-learning-learningprocessandtoprovidesomenewusefulproperties.Manyrealizedexperimentshaveconfirmedtheefficiencyofproposedhybridclustering-classificationneuralnetwork;also,thisnetworkwasusedforsolvingdiagnosticstaskofreactivearthritis.IndexTerms—Hybridclustering-classificationneuralnetwork,supervised/unsupervisedlearning,overlappingclasses,diagnostics,reactivearthritis.I.INTRODUCTIONThecurrentstateoftechnologicaldevelopmentconnectswiththedevelopmentofcomputertools.Thesetoolsdependonthemathematicalmethodsandthepracticalalgorithms.Thedevelopmentofcomputertoolsisactivatorforevolutionofexistingscientificareasandappearanceofnewscientificdirection,forexample,DataScience,BigData,ComputationalIntelligence,DataStreamMiningetc.Moderncapabilitiesofcomputingenvironmentsallowimplementationofalgorithmicallysufficientcomplexmethods,whicharebasisforDataMining.ThehybridneuralnetworksareoneofthemostwidespreadapproachesforsolvingofDataMiningtasks.Self-organizingmaps(SOM)andneuralnetworksoflearningvectorquantization(LVQ)haveseenextensiveuseforsolvingdifferentproblemsinDataMiningdomain(clustering,classification,faultdetectionandcompressionofinformationetc.).ThistypeofneuralnetworkswasproposedbyT.Kohonen[1,2]andrepresents,infact,asingle-layerfeedforwardarchitecture,whichprovidesanoperatorformappingofinputspaceintotheoutputspace.Operation-wiseSOMandLVQarequitesimilartoeachneuronisfedinputsignal(sample)producingoutput,whichisusedduringcompetitionstagetodeterminewinningneuron–usuallytheonewithmaximumoutputsignalvalue.Vectorofsynapticweightsforwinningneuronistheoneclosesttotheinputsampleintermsofthemetricchosen(whichisEuclidianmetricinmostcases).Nextisneuronsadjustmentphase.Synapticweightsofthewinningneurongetsmovedclosertoinputsample.Alternatively,asubsetofneurons(ratherthanasingleone)canbeadjusted–thosedeterminedtobe“reasonablyclose”totheinputsampleareupdated.Resultingnetworkoutput,however,isdeterminedexclusivelybywinningneuron(thisprincipleisusuallyreferredtoas“Winner-Takes-All”(WTA)).Itisthisprinciple(WTA)whichnegativelyaffectsaccuracyincasewhenthereareoverlappingclustersinunderlyingdata.II.RELATEDWORKTakingintoaccounttheabovementionedpropertiesofSOMandLVQnetworks,itmakessensetointroducefuzzyclassificationcapabilitiesontopofthem,whilepreservingonlineoperation.In[8,9]fuzzyself-organizingmapwasproposed,inwhichconventionalneuronsarereplacedbyfuzzyrules.Thisneuralnetworkshowsenoughhighefficiency,butitslearningpropertiesweresignificantlylostespeciallyinon-linemode.2HybridClustering-ClassificationNeuralNetworkintheMedicalDiagnosticsoftheReactiveArthritisCopyright©2016MECSI.J.IntelligentSystemsandApplications,2016,8,1-9In[5,10,11]fuzzyclusteringKohonennetworkandfuzzylinearvectorquantizationnetworkaredescribed.Infact,suchnetworksareneuralrepresentationoffuzzyc-means(FCM)[3],whichisfarenoughfromSOMandLVQmathematicaltoolanddesignedforoperationinbatchmode.III.PROBLEMSTATEMENTLetusconsidersingle-layerneuralnetworkwithlateralconnectionscontainingreceptorsandneuronsintheKohonenlayerwitheachneuronbeingcharacterizedbyavectorofitssynapticweights.Duringlearningstageinputvectorisfedtotheinputsofallneurons(usuallyadaptivelinearassociators)(here-eitherthenumberofobservationinatable“object-properties”,orcurrentdiscretetimeforon-lineprocessingmode)andneuronsproducethescalarsignalsontheiroutputs()()(),1,2,...,Tjjykwkxkjm.(1)Notethatneuron’soutputdependsoncurrentvaluesofsynapticweightsvector,assumingiterativelearningalgorithm.Eachinputvectorcanactivateeithersingleneuron(jw)orasetofneighboringneurons–thisalsodependsonlearningalgorithmchosen.Self-organizationprocedureisbasedonthecompetitivelearningapproach(self-learning)andbeginswiththeinitializationofsynapticweights.Selectinginitialvaluesforweigh