Is there any way to aggregate features from both incoming AND outgoing edges with DGL?

Hi,

I saw the answer in FAQ 1. Are DGLGraphs directed or not? How to represent an undirected graph?
" All DGLGraphs are directed. To represent an undirected graph, you need to create edges for both directions. dgl.to_bidirected can be helpful, which converts a DGLGraph into a new one with edges for both directions."

But in my use case it’s not suitable to do so for memory issues (graph with about one million of edges and a high dimension space for edge feature).

Is there anyway to aggregate features from both incoming AND outgoing edges when updating nodes.

With the message-passing design of DGL, I assume that

g.update_all(fn.copy_e('he', 'm'), fn.sum('m', 'he_aggr'))

will only aggregate features from incoming edges.

It seems that this functionality exists in PyTorch Geometrics (parameter “flow” of message passing method : torch_geometric.nn — pytorch_geometric 2.0.0 documentation ) as in graphnets library (parameters “use_received_edges” and “use_sent_edges” for NodeBlock: graph_nets/blocks.py at master · deepmind/graph_nets · GitHub )

We achieve aggregation of outgoing edges via dgl.reverse(), which should be quite cheap as it reuses the underlying COO/CSR/CSC array. You could perform message passing on the original graph and the reversed graph to get the same goal (which is what flow argument in PyTorch Geometric does).

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.