Design choice: reverse edges and vs undirected edges

[denadai2]

Hi
Let's say I have an hetero network with documentary and tvserie in the graph. I have documentary <has-same-author> tvserie. Let's say I have

A <-> B

where A is 0 and B is 1.

In HeteroData I could have

• edge_index = [[0, 1], [1, 0]]
• edge_index = [[0], [1]] + rev_edges
• edge_index = [[0, 1], [1, 0]] + rev_edges

so I am confused about what's best and what's wrong. Could you help me on this?

thx

[rusty1s]

All edge indices in HeteroData objects needs to point to local node indices (i.e., they need to be in range 0 to num_nodes - 1 of this particular node type). So if you have two node types with a single node each, the resulting edge_index should be defined as

data[tv, doc].edge_index = torch.tensor([[0], [0]])
data[doc, tv].edge_index = torch.tensor([[0], [0]])

[denadai2]

got it, sorry wrong example. However,

should I create rev_edges AND add flipped edges for undirected hetero graphs? (e.g. data[tv, doc].edge_index = torch.tensor([[0, 0], [0, 0]]))

[rusty1s]

No, one of them is enough. You can either create them manually or on the fly via transforms.ToUndirected().

[denadai2]

sry I forgot to reply! Thannkkkss