Skip to content

复现SDNE,并且对领结矩阵部分进行修改,适用大节点

Notifications You must be signed in to change notification settings

shihai-black/SDNE_pytorch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1 Commit
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SDNE(Graph Embedding)

pytorch Based, Structural Deep Network Embedding

修改部分

由于别人都是直接生成拉普拉斯矩阵和领结矩阵,但对于单机器而言,不适用于大规模图数据,因此对领结矩阵和拉普拉斯矩阵部分进行修改。

复现代码

python3 run.py 

参数设置

  -h, --help            show this help message and exit
  --batch_size N        input batch size for training (default: 1024)
  --epochs N            number of epochs to train (default: 5)
  --data S              which data (default: pid)
  -a N, --alpha N       Parameters that control the 1st order loss(default:
                        1e-5)
  -b N, --beta N        The parameters controlling the second-order loss have
                        higher penalty coefficients for non-zero
                        elements(default: 5)
  --v N                 Controls the parameters of the regularization
                        term(default: 1e-5)
  --lr N                Optimizer parameter(default: 1e-3)
  --method METHOD       Classify_method : node classify(n)/link(lp).(default:
                        n)
  -sf N, --sample_frac N
                        Test size
  --cuda                enables CUDA training(default: False)
  --seed S              random seed (default: 1)
  --mode S              The type of graph(default: Di)
  --save S              Whether or not to save(default: y)
  --train S             Train or predict(default: y)
  --module N            Which model to choose(default: SDNE)
  --log-interval N      how many batches to wait before logging training
                        status(default: 1024)

原始数据

参数:

  • 数据集:wiki
  • epoch:5/10/20/50
  • 其他参数均一样
base_epoch5 base_epoch10 base_epoch20 base_epoch50 rec_epoch5 rec_epoch10 rec_epoch20 rec_epoch50
micro_f1 0.4137 0.4927 0.6216 0.5717 0.5841 0.6267 0.7190 0.7072
macro_f2 0.2788 0.3417 0.4479 0.4187 0.3640 0.4208 0.5154 0.5510

小点分析:重构代码效果正常,增加了正则化损失在该数据集更有用

About

复现SDNE,并且对领结矩阵部分进行修改,适用大节点

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages