主動學習以及樣本不均衡在圖數據場景的探索.pdf

《主動學習以及樣本不均衡在圖數據場景的探索.pdf》由會員分享，可在線閱讀，更多相關《主動學習以及樣本不均衡在圖數據場景的探索.pdf（32頁珍藏版）》請在三個皮匠報告上搜索。

1、主動學習與樣本不均衡在圖數據場景的探索周敏 華為云算法創新Lab 主任工程師|自我介紹 本科畢業于中科大，博士畢業于新加坡國立大學 研究方向:圖數據、序列數據模式挖掘和學習01Background0304Conclusion目錄目錄 CONTENTSemantic-aware active learning on graph02Unlabeled Nodes Labeling for imbalanced GraphBackground01|Data available in the form of graphs are ubiquitous.Source from InternetGraph

2、sSocial networksBiology networkAtom networkFinancial networksLogistic networksTelecom networkLink prediction,community detection,node classification,etc.Fraud DetectionGraph Neural Networks are promising tools for fraud detection Label scarce Class imbalanceChallenges in Fraud Detection detectionAct

3、ive Learning on Graphs02|Labels are hard/expensive to collectLabel scarceUnlabelled datalabelled dataPrioritizing the data which needs to be labelled in order to have the highest impact to training a model.Active Learning in Machine LeaningPhoto from interenetPrioritizing the data which needs to be

4、labelled in order to have the highest impact to training a model.Valuable samples-The most informative examples are the ones that the classifier is the least certain about.Active Learning in Machine LeaningPhoto from interenetSelects the most informative nodes as the training labelled nodes based on

5、 the graphical informationDesign different graph-based criteria for node selection on graphs AGE:Uncertainty(entropy)&Representativeness(density¢rality)GRAIN:Influence Maximization&DiversityActive Learning in Graph Machine Leaninghttps:/arxiv.org/pdf/1705.05085.pdfhttps:/arxiv.org/pdf/2108.00219

6、.pdf Mitigating Semantic Confusion from Hostile NeighborhoodSemantic-aware Graph Active Learning Semantic-aware Influence correctionNode influenceSemantic-aware influence1Semantic-aware Graph Active Learning Semantic-aware Influence correctionNode influenceSemantic-aware influence1Semantic-aware Gra

7、ph Active Learning Prototype-based DiversityScore unifyingSemantic-aware Graph Active Learning ExperimentsSemantic-aware Graph Active Learning AnalysisSemantic-aware Graph Active Learning Class Imbalance on Graphs03|Data imbalance leads to decision boundary shift.Imbalance Problem in Machine Leaning

8、Acknowledgement from https:/arxiv.org/pdf/2111.12791.pdfbias induced by imbalanceRe-sampling/re-weighting/cost-sensitivity/hybridSolutions for Learning from Imbalanced DataGraphSmote：Imbalanced Node Classification on Graphs with Graph Neural NetworksFeature extractorSynthetic Node GenerationEdge Gen

9、eratorGNN ClassifierSolutions for Learning from Imbalanced Graph Datahttps:/dl.acm.org/doi/abs/10.1145/3437963.3441720Renode:Topology-imbalance learning for semi-supervised node classificationRe-weight the samples according to their distance to the classification boundaryGraphENS:Neighbor-Aware Ego

10、Network Synthesis for Class-Imbalanced Node Classificationsynthesizes the whole ego network for minor classTAM:Topology-aware margin loss for Class-imbalanced node classificationModify loss based on statistics of the true label distributions of target nodes and classesSolutions for Learning from Imb

11、alanced Graph Datahttps:/ without synthesizing virtual nodes.Take advantage of unlabeled information on graphs.Traditional Self-Training(ST)encounters pseudo-label misjudgement augmentation problem in imbalanced learning.Unlabeled Nodes Retrieval and LabelingDual Pseudo-tag Alignment Mechanism for N

12、ode FilteringNode-ReorderingGeometric ranking Confidence rankingGeometric certain node selectionSelect the most certain node Unlabeled Nodes Retrieval and LabelingExperimentsUnlabeled Nodes Retrieval and LabelingExperimentsUnlabeled Nodes Retrieval and LabelingAcknowledgement華為云算法創新華為云算法創新LabLabAlgo

13、rithm Powers Innovation面向全域，聚焦云智能Top挑戰問題WAFAADAnti-DDoSHSSSASCMCBHVSSCGSDSCDEWDBSSMTDDSCSOC安全VPCELBNAT GatewayElasticIPDirectConnectVPNDNSVPCEPCC網絡計算ECSGACSFACSBMSDCCDeHASIMSCPS應用服務SMNDMSAPMAPIGCPTSCSEROMA FactoryROMA ExchangeBCSAOMAOSCAEECPCROMA ConnectServiceStageEVSSDRSSFSDESOBSDSSDESSCBRCSG存儲Med

14、ia Processing CenterVideo on DemandLiveConvergentVideo Free View VideoVideo Ingestion Service視頻Postgre SQLSQL SeverDDSDDMDRSRDSMySQLDASUGOGaussDB(for Cassandra)GaussDB(for Influx)GaussDB(for Mongo)GaussDB(for MySQL)GaussDB(for openGauss)GaussDB(for Redis)數據庫CCECCISWRAOSMCPIEFASMOSCCIE容器CESDomainsSMN

15、管理云會議網絡研討會會議專業服務智能會議室API&SDK云會議EI企業智能MLSDLSGESCSMRSMTSTTSNLPHiLensDLICloudTableDWSUSearchOCRCDMModerationEIHealthiManufacturing iTransportLogisticsiWaterSmart Retail Smart TrafficImageDLFDISVCTVCMASRNLPImage SearchiPoweriFinanceAISDAPSIQAVision ServiceSmartCampusiRetailImages RecognitionRESDGCTICSCSSLOGIES專屬RegionIEFIEC分布式云HUAWEI CLOUD Stack混合云IoTDAGSLIoT EdgeLiteOSIoTAIoT StageDRIS數字工廠物聯網CodeHubCollaborationProjectManCloudDeployCloudReleaseCodeCICloudPipelineMobileTestDevCloudTestManCodeCheckCloudIDEClassroom軟件開發服務Q&A實驗室目前有多個領域的實習生/博后/研究員崗位開放詳情請關注部門主頁云智能運營/云硬件可靠性管理非常感謝您的觀看|