Prof.ChuanShishichuan@bupt.edu.cnBeijingUniversityofPostsandTelecommunicationsHeterogeneousGraphNeuralNetworkanditsApplicationsinE-Commerce©2009BUPTTSEG223Step-by-StepRepresentationLearning3ØTraditionalrepresentationlearningmodelsusehand-craftedfeaturesandrelativelysimpletrainableclassifier.ØHasthefollowinglimitations:VeryTediousandCostlytodevelophand-craftedfeaturesUsuallyhighlydependentsononeapplication,andcannotbetransferredeasilytootherapplicationshand-craftedfeatureextractor“Simple”TrainableClassifieroutput4End-to-EndRepresentationLearning4ØHierarchyofrepresentationswithincreasinglevelofabstraction.ØImagerecognitionPixel→edge→texton→motif→part→objectØText:Character→word→wordgroup→clause→sentence→storyLow-levelfeaturesoutputMid-levelfeaturesHigh-levelfeaturesTrainableclassifier5Whydonetworkembedding?ØNetworksareagenerallanguagefordescribingandmodelingcomplexsystems5EconomicnetworksSocialnetworksNetworksofneuronsInformationnetworksBiomedicalnetworksInternetABCFigure3:Higher-orderclusterintheC.elegansneuronalnetwork(28).A:The4-node“bi-fan”motif,whichisover-expressedintheneuronalnetworks(1).Intuitively,thismotifdescribesacooperativepropagationofinformationfromthenodesonthelefttothenodesontheright.B:Thebesthigher-orderclusterintheC.elegansfrontalneuronalnetworkbasedonthemotifin(A).Theclustercontainsthreeringmotorneurons(RMEL/V/R;cyan)withmanyoutgoingconnections,servingasthesourceofinformation;sixinnerlabialsensoryneurons(IL2DL/VR/R/DR/VL;orange)withmanyincomingconnections,servingasthedestinationofinformation;andfourURAneurons(purple)actingasintermediaries.TheseRMEneuronshavebeenproposedaspioneersforthenervering(20),whiletheIL2neuronsareknownregulatorsofnictation(21),andthehigher-orderclusterexposestheirorganization.TheclusteralsorevealsthatRIHservesasacriticalintermediaryofinformationprocessing.Thisneuronhasincoming6NetworkEmbeddingNetworkEmbeddinggenerateembedEmbedeachnodeofanetworkintoalow-dimensionalvectorspaceApplication•nodeclassificati...
