Would the data augmentation be label-preserving?

[ha-lins]

Hi @Yuning You,

As the title shows, I have a question about the data augmentation in unsupervised_graph_TU experiments. If we drop an edge or a node and that edge happens to be in a structural motif, it will drastically change the attributes/labels of the molecule. Could you pls give some explainations? Thanks!

[yyou1996]

Hi @ha-lins,

Thanks for your comments. Augmentation represents our prior belief in the data, that what kind of perturbation is rational in the sense of distribution (see Table 1 in paper). Thus, improper augmentation (prior) will lead to deterioration in performance as you imagine, as illustrated in Figure in paper.

[ha-lins]

Thanks for your response. I agree that improper augmentations couldn't be label-preserving and lead to performance degradation.

Besides, have you ever tried different augmentation ratios(except for 10%) on the unsupervised_TU experiments? I think it could be an important hyper-parameter and is about the label-preserving property.

I also find that the GraphCL could not always outperform the Infographmethod over all graph classification datasets significantly. I suggest that more experiments on larger benchmarks such as OGB could be more convincing.

[yyou1996]

We have some OGB results of GraphCL recently in my current project. I will update them gradually. Really appreciate your follow-up discussion.

[ha-lins]

Thanks a lot!