DataFunTalk复杂图迁移学习_2.pdf

《DataFunTalk复杂图迁移学习_2.pdf》由会员分享，可在线阅读，更多相关《DataFunTalk复杂图迁移学习_2.pdf（30页珍藏版）》请在三个皮匠报告上搜索。

1、Knowledge Transfer on Complex Graphs毕文东2024.1.27腾讯TencentA Brief IntroductionCONTENTS Introduction of Graph Knowledge Transfer(GKT)Knowledge Transfer on Graph Data Graph Knowledge Transfer on General DataApplying GKT on both graph and non-graph data2Wendong Bi2Knowledge Transfer on Complex GraphsCON

2、TENTS3Wendong Bi3Knowledge Transfer on Complex Graphs Introduction of Graph Knowledge Transfer(GKT)Knowledge Transfer on Graph Data Graph Knowledge Transfer on General DataApplying GKT on both graph and non-graph datap Graphs are powerful for describing and analyzing entities with relations/interact

3、ions in the real-world data.Webpage NetworksCommon networks in the real world Traffic NetworksBiomedical NetworksKnowledge GraphsWendong Bi4Why GraphsKnowledge Transfer on Complex GraphsData Hungry ProblemWendong Bi5p Data Hungry problem are ubiquitous in the real-worldDeep Learning methods needs ma

4、ssive high-quality dataKnowledge Transfer on Complex GraphsData HungryLow-quality:Web-data with weak annotation,noisy data,etc.Insufficiency:Few training samples for deep learning modelsDeep Learning methods needs massive high-quality dataKnowledge Transfer to Alleviate Data-Hungry Wendong Bi6Knowle

5、dge Transfer on Complex GraphsTransfer valuable knowledge from open-domain data is necessaryIdeal DataOpen Domain DataLow-quality&Large-quantity(Raw data from open domain)High-quality&Small-quantity(Data carefully annotated by humans)Are these large amounts of low-quality data really useless?Connect

6、ed&Transfer*数据来源于IDC数据时代2025Distribution Shift in Graph DataWendong Bi7Node Prediction/ClassificationuvLinkPredictionGraphPrediction/Classificationp Graph Representation Learning aims to learn representations for nodes on the graph,which usually can be divided into three levels(i.e.,Node,Edge,Graph)