NeighborLoader/LinkNeighborLoader on Hetero Data

[colivarese]

Hi, I'm currently trying to work with Heterogeneous Data, the data consists of two node types and ~900 edge types, I'm not sure how to make use of the mini-batch samplers with HeteroData, using the LinkNeighborLoader I have to input the edge_label_index, but I understand it is only possible to pass one edge_type, the same happens with edge_label. When I pass only one edge type it works,

train_loader = LinkNeighborLoader(train_data, num_neighbors = {key: [-1]*2 for key in train_data.edge_types } ,shuffle = True, edge_label_index = train_data.edge_types[0], edge_label=None, batch_size = 128)

The documentation indicates that I can pass a Tensor, EdgeType, or a Tuple with edge types and their indices, but wouldn't it only consider the labels for that particular edge type?

edge_label_index (Tensor or EdgeType or Tuple[EdgeType, Tensor]) – The edge indices for which neighbors are sampled to create mini-batches. If set to None, all edges will be considered. In heterogeneous graphs, needs to be passed as a tuple that holds the edge type and corresponding edge indices. (default: None)

Or do I need to use the NeighborLoader? But I can only pass a node type into input_nodes.

[rusty1s]

This is still TBD, sorry. Currently, only single node types/edge types can be used as seed nodes. We have an open PR out to fix this, see #5013.

[colivarese]

Thanks for the answer, It is good to know. I'll find a way to work it around, have a great day!

[MDrance]

Hello,
I have a question closely related to @colivarese one, I plan on doing link prediction on an hetero graph and something about NeighborLoader and RandomLinkSplit confuses me:

• NeigborLoader doesn't allow us, at the moment, to use multiple nodes types as seed nodes;
• RandomLinkSplit allow me to create train/val/test masks, and from the doc I can read: edge_types (Tuple[EdgeType] or List[EdgeType], optional) – The edge types used for performing edge-level splitting in case of operating on HeteroData objects. (default: None). So I can choose multiple types of edges to create the masks.

If I understood correctly, I could use this to predict links of different types as long as these links start from a node of type specified in NeigborLoader as my seeding. In this case, what would the masks look like ? One for each split considering all edges types or one mask per edge type ?

As an example, would this be correct ?

train_data, val_data, test_data = T.RandomLinkSplit(
    num_val=0.1,
    num_test=0.1,
    neg_sampling_ratio=0.0,
    edge_types=[('user', 'rates', 'movie'), ('user', 'like', 'movie_genre')], #Two edges types
)(data)

loader = HGTLoader(
    hetero_data,
    num_samples={key: [512] * 4 for key in hetero_data.node_types},
    batch_size=128,
    input_nodes=('user', hetero_data['user'].train_mask), #Only one node type as seeding
)

In the case I would like to predict both links ('user', 'rates', 'movie') and ('user', 'like', 'movie_genre') from the only node type allowed in the HGTLoader, here user ?

My final goal is to train a model to predict multiple links of multiples types between multiple types of nodes, maybe there exists an other solution ?

Best,
Martin

[rusty1s]

I don‘t think this is correct. RandomLinkSplit does not create node-level masks - instead it splits edges into message passing edges and edges used for supervision (for each split). As such, you can use RandomLinkSplit to split your dataset for a link-level dataset, and use LinkNeighborLoader to create mini-batches to predict edge-level properties. Currently, we can only handle a single edge type as supervision though (as you correctly pointed out).

[MDrance]

Hello Matthias and thanks for your answer.
I was wondering, is there some place I could find a community to chat with, a slack or maybe a discord of people working with GNN/Pytorch_geometric ?

Best,
Martin

[rusty1s]

Feel free to join our community Slack via https://data.pyg.org/slack.html

[wangjingyamap]

@rusty1s The slack link has expired.

[rusty1s]

Should be fixed.