Aligning negative edge labels generated by RandomLinkSplit with ground-truth negative edge index

[aaronwtr]

Hello PyTorch Geometric community,

I'm currently tackling a heterogeneous link prediction problem and would appreciate some guidance on aligning edge labels with specific edge types. In my dataset, I have two sets of known edge types: positive and negative. The structured representation of my data is in the form of a heterogeneous graph with edge types such as [('A549', 'pos', 'A549'), ('A549', 'neg', 'A549'), ('K562', 'pos', 'K562'), ('K562', 'neg', 'K562'), ('JURKAT', 'pos', 'JURKAT'), ('JURKAT', 'neg', 'JURKAT')]. Here, the central element in each tuple represents the edge type, while the flanking IDs correspond to different cell types (three in total).

I am aiming to perform link prediction like the example provided here. For each specified edge type, I have a distinct edge index dictionary. My challenge lies in ensuring that the edge labels generated by RandomLinkSplit align correctly with the edge indices I've calculated for each edge type. Specifically, I want edges with the 'pos' edge type to have labels set to 1 and edges with the 'neg' edge type to have labels set to 0.

If anyone in the community has insights, advice, or can assist in addressing this challenge, I would greatly appreciate it.

Thank you in advance for your time and support!

Best regards,
Aaron

[aaronwtr]

Fixed it! By using key in RandomLinkSplit, you can point to a self defined attribute. In this case, given I have N indices, I simply labeled them 1 or 0 in a separate variable and passed this to the key parameter in RandomLinkSplit.

[abitofalchemy]

Here you have multi edges (two) per pair of cell type, ea. with pos or neg labels. is that in by design?