lirongwu/krd

Code for ICML 2023 paper "Quantifying the Knowledge in GNNs for Reliable Distillation into MLPs"

Hi there 👋

🌱 I’m currently interested in Graph SSL!

krd's Issues

Reproducing Issue

Hello,

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

While reproducing your code, I observed a large discrepancy in KRD's performance compared to the reported results in a specific setting. On the Citeseer dataset using GAT in a transductive setting, I obtained an accuracy of 55.36 ± 12.69, even though I adopted the same library versions as mentioned. The KRD's performance varied greatly with different random seeds, whereas the teacher GAT maintained stable performance (71.28 ± 1.09).

I believe the method itself is not an issue, as it achieved impressive performance in other settings. However, I'm having trouble pinpointing the exact issue in this particular case.

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

Looking forward to your reply :)