诸神缄默不语-个人CSDN博文目录
本文将对异质图神经网络(HGNN, heterogeneous graph neural networks)的方法演变进行梳理和介绍。
最近更新时间:2023.5.10
最早更新时间:2022.10.31
文章目录
- 1. 异质图
- 2. 处理为同质图
- 3. 知识图谱嵌入
- 4. 传统图学习方法
- 4.1 meta-path系
- 4.2 subgraph系
- 4.3 其他
- 5. GNN + Bi-level aggregation scheme
- 6. GNN + 自监督学习
- 7. 其他
- 8. 其他参考资料
1. 异质图
节点或者边的种类>1即可。
如果节点种类为1,边种类>1,叫multiplex graph。
2. 处理为同质图
直接将节点类型和边类型编码到节点属性中
缺点:不符合GNN内蕴的smoothness假设1,节点/边类型是离散数值,往往与节点特征不共享结构。
3. 知识图谱嵌入
TransE: Translating Embeddings for Modeling Multi-relational Data
DistMult
ComplEx
RotatE
这部分我之前写过相关的博文:cs224w(图机器学习)2021冬季课程学习笔记12 Knowledge Graph Embeddings_诸神缄默不语的博客-CSDN博客
4. 传统图学习方法
4.1 meta-path系
random walk / PageRank
skip-gram
PathSim: Meta Path-Based Top-K Similarity Search in Heterogeneous Information Networks
Meta-Path-Based Ranking with Pseudo Relevance Feedback on Heterogeneous Graph for Citation Recommendation
metapath2vec: Scalable Representation Learning for Heterogeneous Networks:基于metapaths的随机游走+heterogenous skip-gram
HERec: Heterogeneous Information Network Embedding for Recommendation:根据metapath邻居将异质图转换为同质图,用类似DeepWalk的方法学习表征
HIN2Vec: Explore Meta-paths in Heterogeneous Information Networks for Representation Learning:捕获并区别metapaths的信息
多任务学习目标(同时学习节点和metapath的表征)
4.2 subgraph系
将每种关系下的图视作一个子图,然后联合学习这些子图:
Embedding of Embedding (EOE): Joint Embedding for Coupled Heterogeneous Networks
PTE: Predictive Text Embedding through Large-scale Heterogeneous Text Networks:将LINE扩展到异质图上
AspEm: Embedding Learning by Aspects in Heterogeneous Information Networks
4.3 其他
- 嵌入到不同的空间中,或relation embedding:
- HEER: Easing Embedding Learning by Comprehensive Transcription of Heterogeneous Information Networks:考虑异质图中类型不同导致的语义不兼容问题2
改进PTE,通过边表征考虑种类相似性 - PME: Projected Metric Embedding on Heterogeneous Networks for Link Prediction
- HEER: Easing Embedding Learning by Comprehensive Transcription of Heterogeneous Information Networks:考虑异质图中类型不同导致的语义不兼容问题2
- BHIN2vec: Balancing the Type of Relation in Heterogeneous Information Network:随机游走+skip gram+多任务,解决HIN中不同种类边数不平衡的问题
5. GNN + Bi-level aggregation scheme
首先聚合同一类/组中的邻居,然后对其进行聚合(求平均或用注意力机制加权求和)。
RGCN: Modeling Relational Data with Graph Convolutional Networks:每种边一个图卷积层
(2019 NeurIPS) GTN: Graph Transformer Networks:自动学习metapaths(通过图transformer层生成所有可能的联系,在新图上运行图神经网络)
(2019 WWW) HAN: Heterogeneous Graph Attention Network:attentively聚合metapath-based neighborhoods学到的特征
(2019 KDD) HetGNN: Heterogeneous Graph Neural Network:用RWR抽样异质邻居,按节点类型分类,然后聚合
(2020 WWW) MAGNN: Metapath Aggregated Graph Neural Network for Heterogeneous Graph Embedding:首先对节点特征进行转换,然后聚合metapath内部信息,然后聚合各metapath的信息
GATNE: Representation Learning for Attributed Multiplex Heterogeneous Network3
(2020 WWW) HGT: Heterogeneous Graph Transformer:对每个边建模异质attention,隐式学习metapath
缺点:可能会忽略节点信息(尤其在关系种类不平衡时,大类的节点个体信息可能会被忽略 (downweight))
6. GNN + 自监督学习
- 综述
Self-supervised on Graphs: Contrastive, Generative, or Predictive - 研究工作
- (2019) HDGI: Heterogeneous Deep Graph Infomax
- (2020 AAAI) DMGI: Unsupervised Attributed Multiplex Network Embedding:对齐每个视图(通过metapaths生成)上的原始网络和corrupted网络
- (2021 KDD) HeCo: Self-supervised Heterogeneous Graph Neural Network with Co-contrastive Learning:基于metapaths和network分别构建视图
- (2022 SDM) STENCIL: Structure-Enhanced Heterogeneous Graph Contrastive Learning:跨视图+对比学习+结构嵌入
跨视图:保证视图之间一致性最大化→基于metapaths构建视图(metapath实例起终点构成的同质图)→最大化同一节点在不同视图上嵌入的相似性→将各视图的嵌入attentively聚合
结构学习:认为节点嵌入点击不足以建模节点相似性,所以补充了PPR和Laplacian positional embedding
7. 其他
- 一步聚合
- HIME: Heterogeneous graph embedding with single-level aggregation and infomax encoding
对节点特征应用MLP(每一种节点用一个模型),直接进行聚合。损失函数鼓励邻居相近+信息最大化(以促进同质性)
- HIME: Heterogeneous graph embedding with single-level aggregation and infomax encoding
- 考虑不同模态、不同transductive/inductive场景下的GNN场景
inductive链路预测:
DEAL Re9:读论文 DEAL Inductive Link Prediction for Nodes Having Only Attribute Information
LeSICiN Re6:读论文 LeSICiN: A Heterogeneous Graph-based Approach for Automatic Legal Statute Identification fro
这两篇本身做的任务都很niche,但是我认为这个研究topic本身是可以很general、是很有研究价值的(意思是我觉得前途无量,但是我不会做)
8. 其他参考资料
- Heterogeneous Information Network Analysis and Applications
- (2022 Transactions on Big Data) A Survey on Heterogeneous Graph Embedding: Methods, Techniques, Applications and Sources
Revisiting Graph Neural Networks: All We Have is Low-Pass Filters ↩︎
相关阅读笔记博文:
【论文泛读】Easing Embedding Learning by Comprehensive Transcription of Heterogeneous Information Networks_JinyuZ1996的博客-CSDN博客
【论文解读 KDD 2018 | HEER】Easing Embedding Learning by Comprehensive Transcription of HIN_byn12345的博客-CSDN博客 ↩︎相关阅读笔记博文:
GATNE代码部分讲解_酸辣螺丝粉的博客-CSDN博客 ↩︎