Is there any GNN architecture which doesn't use an adjacency matrix internally?

Hi Folks, sorry if my question is too newbie, but all the architectures I’ve already studied until now (mainly convolutional) use an adjacency matrix internally. I was wondering if there is any model/architecture that represents the graph structure in another way.

1 Like

If a model utilizes a graph structure, there is a way to convert the structure into an adjacency matrix, right?

Hi @mufeili. Thanks for your reply. My question is more related to if there are other more performative ways to represent the graph structure than an adjacency matrix. Thus, even being able to generate the adjacency matrix from the graph structure, I was wondering if there is any architecture that actually doesn’t use it directly during training, for instance.