Issue with RandomLinkSplit in train-val split for Heterogeneous Graph

[aaronwtr]

Hi everyone,

I'm encountering unexpected behavior with the RandomLinkSplit in PyTorch Geometric. Before splitting edges, my original graph has 236,448 edges. I'm using the following code for the train-test split:

def train_test_split(graph):
    edge_type_list = list(graph.edge_types)
    rev_edge_type_list = edge_type_list.copy()
    transform = T.RandomLinkSplit(
        num_val=0.15,
        num_test=False,
        is_undirected=True,
        key="target",
        add_negative_train_samples=False,
        edge_types=edge_type_list,
        rev_edge_types=rev_edge_type_list
    )
    return transform(graph)

then calling

train_graph, val_graph, test_graph = random_link_split(graph)

Despite setting num_test to False, the resulting graphs have unexpected edge counts:

• train_graph.num_edges = 200986
• val_graph.num_edges = 200986
• test_graph.num_edges = 236452

Two peculiarities: 1) the test-to-val ratio is not 0.15 as specified, and 2) the test graph contains edges despite explicitly setting num_test to False. The same thing happens if I set num_test = 0.0

Also, if I set num_test = 0.1 and num_val =0.1, I get

• train_graph.num_edges = 189168
• val_graph.num_edges = 189168
• test_graph.num_edges = 212810

I would greatly appreciate any insights into what might be causing this behavior. Thanks in advance for your help!

[rusty1s]

I think you mix up edge_index and edge_label_index. edge_label_index should indeed have a ratio of 0.85/0.15/0. edge_index just denotes the additional edges used for message passing.

[aaronwtr]

I think that makes sense - we are not separating edges in train-test splitting, but rather the edge labels? I suppose I got confused by the docs in RandomLinkSplit which states:

"Performs an edge-level random split into training, validation and test sets of a Data or a HeteroData object (functional name: random_link_split). The split is performed such that the training split does not include edges in validation and test splits; and the validation split does not include edges in the test split."

In this explanation, should it then be edge labels instead of edges?

[rusty1s]

I think the doc string is fine, meaning that no supervision edges of validation and test split are visible for message passing during training; and no supervision edges of the test split are visible for message passing during validation.