search

JiaruiFeng / KP-GNN

Source code for how powerful are K-hop message passing graph neural networks (Neurips 2022)
MIT License
62 stars 5 forks source link

k-hop meaning #3

Closed jw9603 closed 1 year ago

jw9603 commented 1 year ago

hello. I read your paper with interest.

In the K-hop message passing mentioned in this paper, is it correct that L-layer simply refers to the number of repetitions? And in general, in K-hop message passing, we know that layer refers to k-hop. is this correct?

thank you

JiaruiFeng commented 1 year ago

Thanks for your interest in our work! I am glad to answer your questions.

  1. For the first part, yes, the L layer refers to the number of layers in the corresponding GNN model. Or you can consider it as the number of iterations/repetitions of the color update algorithm.
  2. For the second part, I am not sure if I understand your question correctly. Layer in K-hop message passing refers to the same thing mentioned in the first part. Namely, the number of layers in the K-hop message passing GNN model.

If there are any further questions, please let me know!

jw9603 commented 1 year ago

The content of the second question is as follows. As far as I know, the GNN (MPNN) of k layer sees k-hop neighbors, not k iterations

To express it pictorially, it is as follows.

image

In the figure, t is both a t-hop and a t-layer.

Thank you!!

JiaruiFeng commented 1 year ago

Yes, by running k layers of MPNNs, each node can obtain information from its k-hop neighbors. Instead, in k-hop message passing, each node can obtain information from its k-hop neighbors at each layer. This means, by running L layers of k-hop MPNN, each node can get information from L*k-hop neighbors. This is the main difference between the MPNNs and k-hop MPNNs.

Therefore, I think "layer" does not refer to "k-hop" but just the times of message passing in K-hop MPNNs.

jw9603 commented 1 year ago

If so, May i think that there are two types of message passing neural networks?

JiaruiFeng commented 1 year ago

From my perspective, there are many different types of message passing neural networks. The one that people talk about the most is the one discussed in the figure you provided. In our paper, we summarized and analyzed another type of message passing neural network called K-hop message passing. There is a short paper discussed many different message passing neural networks: https://arxiv.org/pdf/2202.11097.pdf

Hope this would help!

jw9603 commented 1 year ago

Thank you for replying even though you are busy!!