zaixizhang / protgnn Goto Github PK

Where can I get the datasets?

Hello, I need the code for a project. I see that the parameters are not exact with respect to the paper. Moreover the node classification functions do not do the prototype projection.

In the paper, the cluster loss is defined as the average minimum of the distance between each graph embedding and the prototype of its own class. However, in the code:

cluster_cost = torch.mean(torch.min(min_distances * prototypes_of_correct_class, dim=1)[0])

The multiplication "min_distances * prototypes_of_correct_class" enables the "distance" of the other class equal to zero, thus the minimum is zero instead of the minimum corresponding distance. As a result, the cluster cost is always zero during the training process, and it doesn't work as expected.

The separation loss has the same problem.

Looking forward to your reply. Thanks!

Hello,
I would like to reproduce the results of the paper related to ProtGNN+. In particular, I am interested in reproducing the results of the reasoning process, similarly to what is reported in Figure 3 of the work. However, in the current version of the repo, I think I am missing how to do it. Could you please provide code/instructions to run ProtGNN+ and visualize the corresponding reasoning process?

Thanks!