tengxiao1/graphacl

GraphACL: Simple and Asymmetric Graph Contrastive Learning (NeurIPS 2023)

graphacl's Issues

Node Clustering Configuration

Thanks for sharing the code!

While attempting to reproduce the node clustering results from the appendix, I utilized the node_embeddings generated by the train function in main.py and conducted kmeans on it. However, I'm facing an issue where the NMI results don't match the reported results.

Could you provide details on any specific settings you applied when evaluating node clustering?

Thanks a lot.

Difficulties encountered in running the code

Hello,

Thank you for your insightful work and for sharing the code.

While reproducing your code with dataset crocodile, I observed that the backpropagation is exceptionally long, resulting in unacceptably slow training. This is also true in the case of datasets for other heterophilic graphs.

I believe the method itself is not an issue. However, I'm having trouble training it.

Could you provide any suggestions that might help resolve this problem?

Looking forward to your reply :)

About one-hop neighborhood context and two-hop monophily

方便起见，我还是用中文吧。
您的工作简洁高效，符合直觉，先👍一个。
论文6.2中有这样的一句：“we attribute our significant improvement to modeling the one-hop neighborhood context and two-hop monophily.” 有一个疑问是这样的：
"two-hop monophily"应该是"one-hop neighborhood context"的补充吧？大多数情况下"two-hop monophily"都会属于"one-hop neighborhood context"的，除了一些例外（比如一个结点只有一个邻居，就没有这种"one-hop neighborhood context"可言）