PyG heterogeneous graph includes the global node id?

[jiaqian]

Hi all,

I want to add the global node id in each batch (on heterogeneous graph) for some post analysis. I know Data object can easily do it by adding data.n_id = torch.arange(data.num_nodes). I tried to add a dictionary by specifying different types of nodes, e.g., Hetero_data.n_id = {"type1":torch.arange(Hetero_data.x_dict["type1"].shape[0]),"type2":...}. In each batch, it has the all the original node ids instead of the ones only in this batch.

Anyone has idea how to do it? Any suggestion is welcome, thanks in advance.

[rusty1s]

For heterogeneous graphs, you would do this per node type:

for node_type in data.node_types:
    data[node_type].n_id = torch.arange(data[node_type].num_nodes)

[jiaqian]

Hi, in my case I work on link prediction task and the label (another type of edge) is also included in edge_index_dict in HeteroData.

1. Does it mean during aggregation process, the label edge information has been used as well?
2. If this is the case, is there any way not using it during aggregation, but I need to use label edge for decoder(take the element-wise product and supply to a fully-connected layer) to train the model in a supervised way.

Thanks in advance.

[rusty1s]

Likely. It is best to ensure that edge_index and edge_label_index do not share information. You can implement this as a simple pre-processing before model training.

[jiaqian]

Thanks for the reply. Is it also possible to use the HeteroConv wrapper and only specify the (graph structure) edge type there rather than label edge? I think that might be easier since I can still use dataloader to batch the label edges for training.

[rusty1s]

You mean to exclude the edge label type as part of the edge types? This should be possible by simply removing this information from the metadata.

[jiaqian]

ok, I see. Thanks.