search

tkipf / gcn

Implementation of Graph Convolutional Networks in TensorFlow
MIT License
7.05k stars 1.99k forks source link

Application of GCN and R-GCN #176

Open AmoghM opened 4 years ago

AmoghM commented 4 years ago

I am new to the GCN research and was trying to differentiate which GCN is suitable for different problems. Here is what I think:

  1. GCN is useful for undirected graph
  2. GCN doesn't utilize edge attributes/features
  3. R-GCN is useful for directed graph
  4. R-GCN utilizes edge attributes/features.

I might be wrong but I really appreciate if you (@tkipf ) could clear this doubt for me. Thanks

QasimWani commented 3 years ago

GCN utilizes edge attributes. previous regularizing techniques consider node similarity, but GCN considers node features. its in the loss function (eq. 8 ). GCN can be used for directed graphs by constructing it as a bipartite graph. See https://www.microsoft.com/en-us/research/wp-content/uploads/2017/01/SDG.pdf for using bipartite graphs for learning on directed graphs