graph4ai/graph4nlp

What is `direction_option`?

smith-co opened this issue · 5 comments

Questions and Help

Couple of questions that I can't figure out neither from documentation nor from Github issues.

  1. Direction_option option could be undirected, bi_sep or bi_fuse. What does it mean really?

  2. Also the documentation is very confusing about relational graph support. It says:

  • heterogenous graph
  • As node

I saw documentation:

There are various dependency relations for dependency graph nodes. According to the need for down-tasks, we provide several options for: 

1. homogeneous graph, 
2. heterogeneous graph. Specifically, for heterogeneous graphs, we support not only various graph edge types but also support bipartite graphs, which regarding the edges as special nodes:

As per this documentation heterogeneous is supported. But then I see some other person asked about heterogeneous graph support and then it was said its not supported! Then I see even RGCN not supported. Super confusing.

  1. What is n_etypes in GGNN? If heterogeneous is not supported then what edge_type really means?
  1. undirected, bi_sep or bi_fuse:
    These options are named according to the message passing. Undierced: undirected graph. Bi-sep: directed graph, the forward and backward information are handled separately, Bi-fuse: directed graph. The forward and backward information are fused together.
    I suggest to read our paper for this library: survey paper.

  2. as_node is the levi-graph which we discussed in our paper. See #529.

  3. n_etypes should be 1 now. This is just a placeholder. Currently it is just a directed homogeneous graph.

Note: the R-GNN is comming soon.

  • Currently all graphs are undirected, then what's the point of bi_sep bi_fuse? It should be set as undirected across all config files. Isn't it?

  • It was mentioned in this issue that RGCN would be supported in March. I looked through all config files, but can't really find out any examples using RGCN. Sincere apologies. Would you mind pointing to the RGCN implementation?

Screen Shot 2022-03-30 at 12 24 37

  1. Sorry. I just want to say we just support homogeneous (directed or undirected) graphs or transformed levi-graph (as_node).
  2. This feature is under developing and there is no stable branch.

@AlanSwift thanks for your response.

For 2, I would not mind even to try out this feature if you give me a working branch and I can do the beta testing 👍

Thanks!

@AlanSwift thanks for your response.

For 2, I would not mind even to try out this feature if you give me a working branch and I can do the beta testing 👍