Dataset with multiple graphs to be used with LightGCN model

[SourabhRanade]

I was trying to make use the lightGCN model that has already been provided under the 'torch_geometric.nn.models' section.
The way I understand it, the model can only take a dataset with one single graph as input.
Is it possible to make use of this model with a dataset that contains multiple small graphs. Something similar to the way PPI dataset is ?

[rusty1s]

LightGCN learns shallow embeddings of nodes, so it is not possible to transfer the results to new graphs. Other than that, I think you could merge your graphs into a single one via Batch.from_data_list(dataset) and go from there.

[SourabhRanade]

Yes, I will start with this!

[rusty1s]

Amazing, thank you :)

[SourabhRanade]

@rusty1s , exams are going on, I have gone through their code ... I'll revert soon with the draft. Sorry for the delay :(

[SourabhRanade]

@rusty1s I was trying to understand the code more in depth from the above mentioned github repository of LighGCN.
Here I was following the program flow for the example dataset lastfm.

Inside the dataloader.py, there is a method getSparseGraph that gets called from the model.py .
Here, they generate the indices, data along with the final sparse graph at the end. I think we can make use of these indexes and data to generate a PyG compatible graph (by small modifications and returning from their code a little early).

I wanted your thoughts on this.

P.S: I am sorry if I have missed out on something obvious. I am new to GNNs and am learning as I explore more.

[rusty1s]

I am not familiar with the LightGCN code base either, but I think we can just provide the datasets similar to how they do it as a torch_geometric.dataset, and then combine model and dataset in a training script (similar to all other examples in PyG).