Question Regarding Self-loop Removal in Graph Structure Learning

[fuyuyuputao]

Hi！While reviewing your code, I noticed that you remove self-loops in the graph structure learning. I am curious about the rationale behind this decision, as I would like to deepen my understanding of the topic. Could you please help me understand why removing self-loops is important or beneficial in Graph Structure Learning?
Thank you for your time and consideration.
Best regards.

[zezhishao]

It will be added in downstream STGNNs. See

STEP/step/step_arch/graphwavenet/model.py

Line 124 in dd702eb

adj_mx = adj_mx + torch.eye(int(adj_mx.shape[1])).unsqueeze(0).expand(B, N, N).to(adj_mx.device)

.