Edge attributes integration and NeighborSampler

[matteomedioli]

Hi there!
I'm working on a multi-relations graph convolutional task in order to obtain node embeddings of Wordnet. I'm using your implementation of GraphSAGE Unsupervised with NeighborSampler. Since for my downstream task, relations between synset or lemma nodes have a strong influence I was wondering if there's a way to include data.edge_attrs(which stores the edge label that identifies the relation type) in SAGEConv (and GCNConv or CATConv too) in the algorithm.

I have two main doubts (sorry in advance for any inaccuracy):

1. Convolutional Layers such as SAGEConv, GCNConv, and GATConv present self.propagate() and self.message() in their forward() method.
   From what I understand, conv layers that manage edge_attrs (GENConv i.e) present in the self.propagate() call the parameter edge_attr (line 152 GENConv).
   Given these two assumptions, can I simply add the edge_attrs in the forward() method for example in SAGEConv, and pass it to self.propagate(), self.message() in order to integrate multi-relations message passing SAGEConv? (add here)

2. Since I can only perform unsupervised training over Wordnet, I was wondering to train the model using negative sampling, as you present in graph_sage_unsup.py using this NeighborSampler implementation. There's a way to manage adge_attrs when sampling neg and pos batch?

Many thanks in advance and thanks for sharing work!

[rusty1s]

1. It depends. If your edge_attr corresponds to an edge type, e.g., holding the relation type of an edge as a LongTensor, I suggest to look at the RGCNConv, where you apply different SAGEConv modules for each distinct edge type. Otherwise, yes, you can simply concatenate or add the edge_attr to the neighboring node features in a customized SAGEConv implementation.
2. You need to sample or generate random edge_attr values as well. If that is just an edge type, you can simply sample any type from the number of relations.