How to generate datasets for inductive learning?

[scottshufe]

Hi, everyone. My question is how to generate datasets for inductive learning?
Using Cora dataset as an example, I want to use half of its nodes and their links for training, the rest for testing. I am not sure if this approach is correct:

1. randomly sample half of nodes on the graph
2. use the sampled nodes to generate subgraph(1) with torch_geometric.utils.subgraph function, the rest nodes for subgraph(2)
3. use the generated subgraph(1) as the training dataset and subgraph(2) for testing

Does anyone have better ideas?

[wsad1]

For datasets like cora I have mostly only seen examples of transductive learning (nodes in same graph are split into train/test. although this is not completely transductive as there is nothing stopping you from predicting on an unseen node) as done here.

The approach you are suggesting sounds interesting, but because you are generating subgraphs, the neighborhood of a node might change, which might effect your prediction.

[scottshufe]

Got it. To make it simple, I just want to split a graph into two parts.
Thanks for your suggestion. :)

[rusty1s]

train_mask = torch.rand(data.num_nodes) < 0.5
test_mask = ~train_mask

train_data = copy.copy(data)
train_data.edge_index, _ = subgraph(train_mask, data.edge_index, relabel_nodes=True)
train_data.x = data.x[train_mask]

test_data = copy.copy(data)
test_data.edge_index, _ = subgraph(test_mask, data.edge_index, relabel_nodes=True)
test_data.x = data.x[test_mask]

[scottshufe]

Thank you Matthias! I think this is what I need.