Mismatched Edges in HGTLoader for OGB_MAG Dataset

[Kattasme]

Hey

I've been experimenting with PyTorch Geometric's HGTLoader to train a HGT model on the OGB_MAG dataset. Everything seems to be running smoothly, but upon double-checking the HGTLoader, I discovered that the edges for type X lets say ('paper', 'cites', 'paper') don't match with the original dataset. This shouldn't be happening, right ? What could be causing this issue? Or do I miss something?

Here some sample code

    # types
    main_node = 'paper'
    main_edge_type = ('paper', 'cites', 'paper')

    # Loading dataset
    path = osp.join(osp.dirname(osp.realpath(__file__)), path_)
    dataset = OGB_MAG(root=path, preprocess='metapath2vec', transform=T.ToUndirected())
    data = dataset[0]
    data = data.edge_type_subgraph([main_edge_type])
    data = T.RemoveIsolatedNodes()(data)

    # Create HGTLoader
    train_loader = HGTLoader(
        data,
        num_samples={key: [5] * 2 for key in data.node_types},
        batch_size=10,
        input_nodes='paper',
    )
    batch = next(iter(loader))

If we now compare the batch edge index

`batch[main_edge_type].edge_index`

with the original data edge index

`data[main_edge_type].edge_index`

we can see, that some, or all edges from the batch do not exist in the original data.

I am very grateful for any help.

[Kattasme]

I will answer myself.

The difference in edge indices between the original dataset (data[main_edge_type].edge_index) and the batch (batch[main_edge_type].edge_index) is due to the sampling process of the HGTLoader. The loader samples subgraphs, which leads to reindexing of nodes and edges within each batch.