Why is the edge from dst to src in DGL graph? about magnn HOT 4 CLOSED

For example, if we have a graph with 3 nodes [0, 1, 2] and edges [(0, 1), (0, 2), (1, 0), (2, 0)]
If we sample node 0 as the target node for graph convolution, parse_minibatch would firstly fetch the adjlist of node 0, that is [(0, 1), (0, 2)]
In DGL, the natural way to do graph convolution is to let message flow from the source nodes to the target node. So here I reverse the node pairs to be [(1, 0), (2, 0)] to let message flow from node 1 and node 2 to node 0.

from magnn.

Thanks for your reply! I think I understand.

But I still want to confirm whether the basis of this operation is that the relationship between the two nodes of DBLP is always bidirectional. (e.g., write vs. written by)

For example, in the graph [(0, 1), (0, 2), (1, 0)], the relationship between nodes is not always bidirectional, so when updating the target node 0, the extracted edge should be [(1,0)] instead of [(1,0),(2,0)].

Is this understanding correct?

from magnn.

Yes, in my preprocessing and model, each relationship between two nodes is converted to two directional edges. E.g., "A writes B" would be treated as "A writes B" and "B is written by A".

from magnn.

ok, thank you very much!

from magnn.