"to_bidirected is not well defined for unidirectional bipartite graphs" when required for GNN

[kunal-bhadra]

Title. I am trying to convert my simple graph to test link prediction, and I come across the same issue quoted here. When I go to use the to_bidirected I get the below error.

to_bidirected is not well defined for unidirectional bipartite graphs, but (‘customer’, ‘customerBoughtProducts’, ‘product’) is unidirectional bipartite

I am not sure how to get past this, since I am testing this bipartite graph but also need the message passing to happen correctly.

Edit: Graph Dimensions

Graph(num_nodes={‘customer’: 74845, ‘product’: 1835},
num_edges={(‘customer’, ‘customerBoughtProducts’, ‘product’): 266617},
metagraph=[(‘customer’, ‘product’, ‘customerBoughtProducts’)])

[Rhett-Ying]

you wanna add reverse edges of (‘customer’, ‘customerBoughtProducts’, ‘product’) ?
Such a new graph is still a bipartite graph?
How about constructing a new graph via dgl.heterograph.

[kunal-bhadra]

Yes, I did just that. Thought it’d be easier to just add another edge with the tensors in reverse positions and it worked.