Given a set of source node indices and destination node indices, what is the difference between constructing a block vs. a subgraph? (Other than the fact that a subgraph is not a bipartite heterogenous graph).
If destination node indices are a subset of source node indices, would a subgraph also suffice in minibatch training as opposed to blocks?
You can use subgraphs in minibatch training and there are such papers like ClusterGCN and GraphSAINT. In the meantime, you may view blocks more like computational graphs for describing the computational dependencies. In a DGLBlock, there can be two copies of the same node in both source and destination nodes.
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.