知识表示学习方法调研

NeuralKG

https://github.com/zjukg/NeuralKG

NeuralKG: An Open Source Library for Diverse Representation Learning of Knowledge Graphs

ComponentsModels
KGEModelTransE(2013), TransH(2014), TransR(2015), ComplEx(2016), DistMult(2014), RotatE(2019), ConvE(2018), BoxE(2020), CrossE(2019), SimplE(2018)
GNNModelRGCN(2017), KBAT(2019), CompGCN(2020), XTransE(2020)
RuleModelComplEx-NNE+AER(2018), RUGE(2018), IterE(2019)

知识表示学习常用的方法

来自第四范式技术分享

https://yeu.h5.xeknow.com/sl/3BzAvx

https://github.com/AutoML-Research

Triplet-based models

Translational Distance Model(TDM)

2013,TransE

2014,TransH

2019,TotatE

Neural Network Model(NNM)

2014,MLP

2018,ConvE

2020,RSN

BiLinear Model(BLM)

2011,RESCAL,关系为方阵

2014,DistMult,关系为对角矩阵

2016,CompIEx,

2017,Analogy

2018,SimpIE

2019,QuatE

2020,AutoSF,BLMs – unified form

2021,AutoSF+,An AutoML approach to design bilinear scoring functions

Bilinear models are the best choice in triplet-based models

Relational path

2015,PTransE

2019,RSN,Models long-term information well, but is limited in modeling short-term information

2020,Insterstellar,A NAS approach to recurrently process the relational paths

The relational paths contain richer information than triplets

GCN-variants

2017,R-GCN

2020,CompGCN

2020,GraIL,Enclosing subgraph -> Entity labeling – > GNN scoring

2021,KE-GCN

2022,RED-GNN

发表回复

您的电子邮箱地址不会被公开。 必填项已用 * 标注