How does RandomNodeSplit and RandomLinkSplit work together?

[hash-ir]

Hi,

I have created train, val, and test masks using RandomNodeSplit. Then, I create the associated loaders using NeighborLoader. Since I am using autoencoder architectures, I need to split the edges into pos and neg edge indices. I found the RandomLinkSplit but I don't know how it works with the node split already done. Do I apply this transformation on the full dataset before the node split or to the train, val, and test batches separately during training? In the latter case, it would again divide, for example, a training batch, into pos and neg train, val, and test edge indices that do not make sense.

Basically, I want to avoid a data leak in terms of the node or edge indices between the train, val, and test sets.

Thanks,
Hashir

[rusty1s]

Can you say a bit more about what it means to apply RandomNodeSplit and RandomLinkSplit together? What do you want to try to achieve? What do you expect to happen?

One thing I could think of is to split the graph into 3 subgraphs via data.subgraph() based on the node-level masks coming from RandomNodeSplit. Each subgraph can then be further splitted via RandomLinkSplit.

[hash-ir]

I want to split the dataset based on nodes and get positive and negative edge labels and indices for the autoencoder. What I notice is that RandomLinkSplit also divides the data object into train, val, and test. This is where I am a bit confused.

If I split the graph using data.subgraph(), the resulting subgraphs will then again be divided into three sets via RandomLinkSplit.

[rusty1s]

Yeah, you can probably omit the use of RandomLinkSplit, and treat the isolated subgraphs obtained from RandomNodeSplit as training, validation and test graphs for link prediction.