论文网址:[1611.07308] Variational Graph Auto-Encoders (arxiv.org)
英文是纯手打的!论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误,若有发现欢迎评论指正!文章偏向于笔记,谨慎食用
目录
2.1. A latent variable model for graph-structured data
2.2. Experiments on link prediction
1. 省流版
1.1. 心得
(1)好短的文章捏,只有两页
1.2. 论文总结图
2. 论文逐段精读
2.1. A latent variable model for graph-structured data
①Task: unsupervised learning
②Latent space of unsupervised VGAE in Cora, a citation network dataset:
③Definitions: for undirected and unweighted graph , the number of nodes
, the adjacency matrix with self-loop and the diagnal elements all set to 1defined as
, the degree matrix is
, the stochastic latent variables is
,
, node feature matrix
(但是没说这个节点特征是啥,估计自己随便定义吧)
④Inference model:
with
where is the matrix of mean vectors
;
(为啥左边要有个log啊)
⑤A 2 layer GCN:
where denotes weight matrix,
⑥ 和
共享
的参数???什么玩意儿??为啥有俩,是引用了之前的什么高斯吗?
⑦Generative model:
with
where represents the logistic sigmoid function
⑧Loss function:
where Gaussian prior
⑨作者觉得对于非常稀疏的邻接矩阵,在损失函数中重新加权a)
的项,或b)
的子样本项可能是有益的。然后它们选择了a) 方法。
⑩If there is no node features, replace by indentity matrix
⑪Reconstruct adjacency matrix by non-probabilistic graph auto-encoder (GAE) model:
2.2. Experiments on link prediction
①Prediction task: randomly delete some edges and keep all the node features
②Validation/Test set: deleted edges and unconnected node pairs with the same number
③Connection contained: 5% for val set and 10% for test set
④Epoch: 200
⑤Optimizer: Adam
⑥Learning rate: 0.01
⑦Hidden dim: 32
⑧Latent variable dim: 16
⑨Embedding dim: 128
⑩Performance comparison table with mean results and std error for 10 runs:
where * means w/o node features
3. Reference
Kipf, T. N. & Welling, M. (2016) 'Variational Graph Auto-Encoders', NIPS. doi: https://doi.org/10.48550/arXiv.1611.07308