第十一篇论文讲解.pptxVIP免费

贪心科技|让每个人享受个性化教育服务Review:PaperReadingGraphTransformerNetworks范老师2020/06/18贪心科技|让每个人享受个性化教育服务Content•Introduction•RelatedWork•Method•Experiments•ConclusionandFutureWork贪心科技|让每个人享受个性化教育服务Background贪心科技|让每个人享受个性化教育服务Background贪心科技|让每个人享受个性化教育服务Introduction•GraphNeuralNetworks(GNNs)havebeenwidelyadoptedinvarioustasksovergraphs,suchasgraphclassification,linkpredictionandnodeclassification.•Homogeneousgraph:astandardgraphwithonetypeofnodesandedges•Heterogeneousgraph:agraphwithmorethanonetypeofnodesandedges•WedevelopGraphTransformerNetwork(GTN)thatlearnstotransformaheterogeneousinputgraphintousefulmeta-pathgraphsforeachtaskandlearnnoderepresentationonthegraphsinanend-to-endfashion.贪心科技|让每个人享受个性化教育服务RelatedWorkGraphNeuralNetworks包括spectral和non-spectral两种方法。spectral是基于spectraldomain(使用傅里叶变换)上进行卷积。non-spectral直接在图上沿着edge对空间特征进行卷积。NodeclassificationwithGNNsRelatedwork只在同构图上做计算，异构图的计算需要先建立元路径，转化为同构图在做计算都需要手工构建元路径。这些元路径的选择对下游算法的准确性有很大的影响。贪心科技|让每个人享受个性化教育服务Method-Preliminaries贪心科技|让每个人享受个性化教育服务Method-Preliminaries123110000000000000000000000010edge2edge1只含有edge1的邻接矩阵只含有edge2的邻接矩阵含有edge1+edge2的metapath的邻接矩阵贪心科技|让每个人享受个性化教育服务Method-Preliminaries贪心科技|让每个人享受个性化教育服务Method–Meta-PathGeneration贪心科技|让每个人享受个性化教育服务Method–GraphTransformerNetworks贪心科技|让每个人享受个性化教育服务Experiments–SettingsandBaselines基于局部信息的图算法：DeepWalk,metapath2vec基于全局信息的图算法：GCN,GAN,HAN贪心科技|让每个人享受个性化教育服务Experiments–ResultsonNodeClassification1.OurGTNmodelachievedthebestperformancecomparedtoallotherbaselinesonallthedatasetseventhoughtheGTNmodelusesonlyoneGCNlayerwhereasGCN,GATandHANuseatleasttwolayers.ItdemonstratesthattheGTNcanlear...