Message passing for directed graphs

Hi, thanks for the great package. I find some potential unclarity of the message passing mechanism when applied to directed graphs. In particular, the edge_index includes pairs of (i, j) in the edge set, but the aggregation associated with the MessagePassing class works by aggregating features from x_j to x_i, i.e. from target nodes to source nodes. This is fine for undirected graphs, but for directed graphs, we actually want to aggregate features flowing into a node (usually not the opposite way), so x_j should be the “source node” and x_i should be the “target node”. I find these terminologies in the documentation about “source” and “target” a bit misleading.

Perhaps if not changing the aggregation direction in the MessagePassing code, in the definition of edge_index, we should make it clear that a pair (i, j) in edge_index means that there is an edge from j to i, i.e. in the direction of j -> i. This does not influence undirected graphs but helps to reduce errors in dealing with directed graphs.

[rusty1s]

Added to current master!

[jzhou316]

Both are fine. 1 is perhaps more systematic? Seems can do a minimal change by modifying edge_index to edge_index[torch.LongTensor([1, 0])] in MessagePassing.propagate() according to the direction flag.