Data integration using GNNs

[paulamartingonzalez]

I am using GNNs for the graph classification setting (i.e. one graph with their nodes and node features, one label). I have some extra features that I think may be relevant to predict the label, they are characteristics of the whole graph (i.e. one vector per graph).

I am considering two options to do this integration: (1) concatenate them in the latent space - where the GNN gets a feature vector per graph or (2) use them as node features - all the same per graph. Is there any intuition of what would be the best way to do this?

Explainability of the results is very important, would I be able to do something like this using GNNExplainer if I use setting (1)?

Thanks in advance!

[mufeili]

If I were you, I will start with (1). Specifically, concatenate the graph-level representation with the input graph feature, potentially projected by a linear transformation/MLP. (2) might also be feasible, but will be more costly. The benefit is that it allows interaction with input node feature and input graph features in GNN layers, which might further improve the model performance.

Either way you can still apply a GNNExplainer. It’s more about what kind of explanation you are seeking.