Hi there.
I had a question related to adding some time based dependency on the graph where we can have 2d convolution over multiple graphs differing in features/node. An approach taken is Spatio-Temporal GNN (STGNN) where they combine RNN + GCN. But they have more dynamic graphs (correct me if I'm wrong).
Is there a way to alleviate GNN formulation to compute 2d convolutions as well? I guess it should be straightforward, no? Instead of 1 adjacency matrix as input, we take in a batch and compute the convolution (eq. 7 of GCN paper) together. Would that not work?
Hi there. I had a question related to adding some time based dependency on the graph where we can have 2d convolution over multiple graphs differing in features/node. An approach taken is Spatio-Temporal GNN (STGNN) where they combine RNN + GCN. But they have more dynamic graphs (correct me if I'm wrong). Is there a way to alleviate GNN formulation to compute 2d convolutions as well? I guess it should be straightforward, no? Instead of 1 adjacency matrix as input, we take in a batch and compute the convolution (eq. 7 of GCN paper) together. Would that not work?
Thanks