当前位置:网站首页>Variational graph auto-encoders (VGAE)
Variational graph auto-encoders (VGAE)
2022-06-28 19:07:00 【连理o】
Contents
Variational graph auto-encoders (VGAE)
Graph Auto-Encoders (GAE)
Definitions
- 给定一个无向无权图 G = ( V , E ) \mathcal G=(\mathcal V,\mathcal E) G=(V,E), N = ∣ V ∣ N=|\mathcal V| N=∣V∣ 为顶点数, A ∈ R N × N \boldsymbol A\in\R^{N\times N} A∈RN×N 为邻接矩阵 (对角线元素为 1), X ∈ R N × D \boldsymbol X\in\R^{N\times D} X∈RN×D 为结点的特征向量
Graph Auto-Encoders

- GAE 中的 Encoder 为 GCN,它负责由邻接矩阵和结点特征编码得到每个结点的 embedding 向量 z i z_i zi ( i = 1 , . . . , N i=1,...,N i=1,...,N),它们构成了结点 embedding 矩阵 Z ∈ R N × F \boldsymbol Z\in\R^{N\times F} Z∈RN×F
- GAE 中的 Decoder 为一个简单的 inner product decoder,它负责由结点的 embedding 向量 Z \boldsymbol Z Z 来重构邻接矩阵 A ^ \hat \boldsymbol A A^。它通过计算 σ ( z i T z j ) \sigma(z_i^Tz_j) σ(ziTzj) 来决定 A ^ i j \hat \boldsymbol A_{ij} A^ij

A framework for unsupervised learning on graph-structured data
- GAE 引入自编码器来处理图数据,可以基于图数据进行无监督学习
Variational graph auto-encoders (VGAE)
- VGAE 在 GAE 的基础上进一步引入了 变分自编码器 (VAE) 的思想,对 latent space 施加正则化来保证一个 regular latent space

- VGAE 假设先验概率服从标准正态分布
p ( Z ) = ∏ i p ( z i ) = ∏ i N ( z i ∣ 0 , I ) p(\boldsymbol Z)=\prod_i p(z_i)=\prod_i \mathcal N(z_i|0,\boldsymbol I) p(Z)=i∏p(zi)=i∏N(zi∣0,I) - 似然由点积模型得到

- 后验概率由变分推理近似得到,概率分布族为协方差矩阵为对角矩阵的正态分布
其中, μ = GCN μ ( X , A ) \mu=\text{GCN}_\mu(\boldsymbol X,\boldsymbol A) μ=GCNμ(X,A) 为 GCN 输出的后验概率分布均值向量, log σ = GCN σ ( X , A ) \log\sigma=\text{GCN}_\sigma(\boldsymbol X,\boldsymbol A) logσ=GCNσ(X,A) 为 GCN 输出的后验概率分布标准差的对数。 GCN \text{GCN} GCN 为一个简单的 2 层 GCN,可以被表示为 GCN ( X , A ) = A ~ RELU ( A ~ X W 0 ) W 1 \text{GCN}(\boldsymbol X,\boldsymbol A)=\tilde \boldsymbol A\text{RELU}(\tilde \boldsymbol A\boldsymbol X\boldsymbol W_0)\boldsymbol W_1 GCN(X,A)=A~RELU(A~XW0)W1,其中 W i \boldsymbol W_i Wi 为 MLP 权重矩阵, A ~ = D − 1 2 A D − 1 2 \tilde \boldsymbol A=\boldsymbol D^{-\frac{1}{2}}\boldsymbol A\boldsymbol D^{-\frac{1}{2}} A~=D−21AD−21 为归一化的邻接矩阵, D \boldsymbol D D 为度矩阵 (一个对角矩阵,对角元素为各个顶点的度数),左乘 D − 1 2 \boldsymbol D^{-\frac{1}{2}} D−21 会使得 A \boldsymbol A A 的第 i i i 行除以结点 i i i 度数的根号,右乘 D − 1 2 \boldsymbol D^{-\frac{1}{2}} D−21 会使得 A \boldsymbol A A 的第 i i i 列除以结点 i i i 度数的根号,因此 A ~ i j = A i j / ( D i i D j j ) \tilde\boldsymbol A_{ij}=\boldsymbol A_{ij}/(\sqrt{\boldsymbol D_{ii}\boldsymbol D_{jj}}) A~ij=Aij/(DiiDjj),相当于是给邻接矩阵根据度数做了一个归一化。 A ~ X = [ a ~ 1 T X . . . a ~ N T X ] \tilde \boldsymbol A\boldsymbol X=\begin{bmatrix} \tilde a_1^T\boldsymbol X\\...\\\tilde a_N^T\boldsymbol X\end{bmatrix} A~X=⎣⎡a~1TX...a~NTX⎦⎤ 是在进行结点的信息聚合。 GCN μ ( X , A ) \text{GCN}_\mu(\boldsymbol X,\boldsymbol A) GCNμ(X,A) 和 GCN σ ( X , A ) \text{GCN}_\sigma(\boldsymbol X,\boldsymbol A) GCNσ(X,A) 共享第一层的权重 W 0 \boldsymbol W_0 W0 - 由变分推理得到的优化问题为最大化下式:

References
- Kipf, Thomas N., and Max Welling. “Variational graph auto-encoders.” arXiv preprint arXiv:1611.07308 (2016).
- Wu, Zonghan, et al. “A comprehensive survey on graph neural networks.” IEEE transactions on neural networks and learning systems 32.1 (2020): 4-24.
- code: https://github.com/DaehanKim/vgae_pytorch
边栏推荐
- math_ Proving common equivalent infinitesimal & Case & substitution
- 180.1. Log in continuously for n days (database)
- leetcode 1689. Partitioning into minimum number of deci binary numbers
- 智能计算系统2 bangc算子开发的demo (CPU和MLU270的异构编程流程)
- C# 41. int与string互转
- 3D可旋转粒子矩阵
- I just bought the ADB MySQL service. Every time I do an operation, such as creating a table, this problem will pop up. What is the problem?
- Cvpr2022 | Zhejiang University and ant group put forward a hierarchical residual multi granularity classification network based on label relation tree to model hierarchical knowledge among multi granu
- use. NETCORE's own background job, which simply simulates producers and consumers' processing of request response data in and out of the queue
- MDM数据分析功能说明
猜你喜欢

1 goal, 3 fields, 6 factors and 9 links of digital transformation

Idea merge other branches into dev branch

MDM数据分析功能说明

Cvpr2022 | Zhejiang University and ant group put forward a hierarchical residual multi granularity classification network based on label relation tree to model hierarchical knowledge among multi granu

基于管线的混合渲染

Hands on Teaching of servlet use (1)

How to change the status bar at the bottom of win11 to black? How to change the status bar at the bottom of win11 to black

Mybayis之核心主件分析

深度学习需要多强的数学基础?

Markdown绘图mermaid实用教程
随机推荐
行业分析| 快对讲,楼宇对讲
High performance and high availability computing architecture scheme commented by Weibo
Web3 that unleashes the value of the Internet
An in-depth analysis of the election mechanism in kubernetes
Anonymous function variable problem
Jenkins Pipeline 对Job参数的处理
泰山OFFICE技术讲座:WORD奇怪的字体高度
展示用sql创建中间数据表的实际工作用例
Win11如何给系统盘瘦身?Win11系统盘瘦身方法
Opengauss kernel: analysis of SQL parsing process
应用实践 | 10 亿数据秒级关联,货拉拉基于 Apache Doris 的 OLAP 体系演进(附 PPT 下载)
pd.cut 区间参数设定之前后区别
Leetcode 周赛299
CVPR2022 | 浙大、蚂蚁集团提出基于标签关系树的层级残差多粒度分类网络,建模多粒度标签间的层级知识
大火的虚拟人在哪些产业开始发力?
How does the computer check whether the driver is normal
Idea merge other branches into dev branch
智能计算系统3 Plugin 集成开发的demo
ANR Application Not Responding
try except 添加辅助新列