Representing Bipartite graphs with concatenated features

[ebrahimpichka]

Hello PyG team/community,

I have issues with working with bipartite graphs that i'd appreciate your help with.

Suppose we have two separate sets of nodes which form our Undirected Bipartite graph i.e. $V = ( V_{src}, V_{dst} ) $.

The docs mention that

For bipartite graphs with two independent sets of nodes and indices, and each set holding its own information, this split can be marked by passing the information as a tuple, e.g. x=(x_N, x_M)

First question

My first question is, would it make a difference if we just construct one graph with (x_N, x_M) stacked together and make sure that node indices of dst starts after that of src. Like:

num_src_nodes = len(x_src)
num_dst_nodes = len(x_dst)

src_indices = torch.tensor([src for src in range(num_src_nodes) for _ in range(num_dst_nodes)])
dst_indices = torch.tensor([dst for _ in range(num_src_nodes) for dst in range(num_dst_nodes)])

# Shift dst indices to ensure they're distinct from src
# edge_index with shape [2, num_edges]
edge_index = torch.stack([src_indices, dst_indices + num_src_nodes])
    
# single feature mat
x = torch.cat([x_src, x_dst], dim=0)
    
data = Data(
    x=x,
    edge_index=edge_index,
    )

Second question

If there is indeed a difference, what's the advantage of separating src and dst for conv layers?
Also, is there an example of how using the x=(x_N, x_M) method works? i mean the entire steps of preparing the data, defining the network, propagation, etc.