几个综述:
- Graph Embedding Techniques, Applications, and Performance: A Survey
- Representation Learning on Graphs: Methods and Applications
- A Comprehensive Survey of Graph Embedding:Problems,Techniques and Applications
网络表达的任务
- reconstruction
- Network Reconstruction:
学习network表达–>预测existing link(nearest neighbor)
- inference
- node classification
学习network表达–>训练分类器–>分类数据 - link prediction
学习network表达–>预测feature link(nearest neighbor)
- insight
- visualization
学习network表达–>layout
浅层Network embedding
node&link
–>
node neighborhood:
DeepWork(random work+ skip gram)
Node2Vec(BFS&DFS ---- next node with probability p&q https://www.jianshu.com/p/c631500d57dd)
–>
pair-wise proximity:
LINE:
first order proximity 最小化 两个顶点的向量表示的相似度(联合概率分布) 和 两个顶点的边的权重(经验) 的 KL散度;
second order proximity 最小化 两个顶点的相似度在当前节点与所有邻居相似度的比重(联合概率分布)和两个顶点的权重在当前点相连的边的占比
https://www.jianshu.com/p/90784081bc70
–>
community structure
M-NMF(Community Preserving Network Embedding):
microscopic structure和 community structure的结合
https://zhuanlan.zhihu.com/p/33405615?from_voters_page=true
Structural Deep Network Embedding(SDNE)
图拉普拉斯的理解
https://www.cnblogs.com/xbinworld/archive/2012/11/29/2795287.html
SDNE方法其实就是Auto-encoder为网络,误差为Auto-encoder输入和输出与first-order
https://blog.csdn.net/wangyang163wy/article/details/79698563
Property-preserving network embedding
-
Non-Transitivity:
Non-transitive Hashing with Latent Similarity Components
论文的方法详见我之前的博客。 -
Asymmetric Transitivity
Asymmetric Transitivity Preserving Graph Embedding
Network embedding with side information
-
Node Content
Network Representation Learning with Rich Text Information
http://www.rootlu.com/blog/2017/08/07/TextDeepWalk.html/ (这个小哥自己做了一个网站,在上面写一些博客)
https://blog.csdn.net/zhangbaoanhadoop/article/details/82313123 -
Node Label
Max-Margin DeepWalk: Discriminative Learning of Network Representation -
Node Content + Label
Tri-Party Deep Network Representation
https://blog.csdn.net/NockinOnHeavensDoor/article/details/80687330
Heterogeneous Information Network 异构信息网络做网络嵌入
- meta-path
最早提出meta-path的是sunyizhou:
Meta Path-Based Top-K Similarity Search in Heterogeneous Information Networks. - Heterogeneous Information Network Embedding for Meta Path based Proximity.
- Heterogeneous Network Embedding via Deep Architectures
- Comparing Apples to Oranges: A Scalable Solution with Heterogeneous Hashing
Application-oriented network embedding
- Cascade prediction
- anchor link prediction
Future Directions
-
Dynamic Networks
动态网络如何做embedding,动态网络的结构会变,比如增加/删除点/边。
TIMERS: Error-Bounded SVD Restart on Dynamic Networks
Learning Embeddings of Out-of-Sample Nodes in Dynamic Networks -
Hyper-Networks
Structural Deep Embedding for Hyper-Networks -
More Complex Structure
-
Preserve Special Property
-
Role of Side information
-
Embedding Space
-
End-to-End solution
-
Scalability