Two questions of the implements of R-GCN

Questions
#1

Dear authors,
Thanks for reading my questions,

  1. You have a new batch sampling method in the lastest version of your R-GCN code. Specifically,
    Sample edges by neighborhool expansion and Sample edges uniformly from all the edges. Dose
    the results of the R-GCN influence by the two batch sampling method heavily?
  2. In the lastest studies of GCN batch sampling methods, there are two kinds of methods: graph
    sampling and layer sampling. Why all relation-aware GCNs use the graph sampling rather than
    the layer sampling??
  3. The ReadME of the lastest version of R-GCN do not match the code, there is no heperparameter called ‘–filter’ or ‘–raw’ in your code

Look forward for your reply!

#2

In this PR: https://github.com/dmlc/dgl/pull/1516, we use layer sampling.
And we found graph sampling is good for link prediction tasks.