search

LFhase / GALA

[NeurIPS 2023] Does Invariant Graph Learning via Environment Augmentation Learn Invariance?
MIT License
14 stars 1 forks source link

About the differences between the results of two papers. #1

Open qkrdmsghk opened 8 months ago

qkrdmsghk commented 8 months ago

Hi LFhase, thank you for your contributions to the GALA method.

I recently found two versions of the GALA publications. One is https://openreview.net/forum?id=bjw5jqGtDy which was published at ICLR 2023 Workshop on Domain Generalization. The other is https://openreview.net/forum?id=EqpR9Vtt13 which was published at NeurIPS 2023.

Both papers proposed GALA methods, but their experimental results on DrugOOD datasets are inconsistent. I am wondering the reason of the difference, and which version is the benchmark for comparison?

Thanks for taking the time to read this! Your attention means a lot to me.

LFhase commented 8 months ago

Hi @qkrdmsghk , thanks for noting the differences in the results. In the updated results published at NeurIPS, we adopted a better backbone for the subgraph extraction module, where we incorporated InstanceNorm into the edge attention module inspired by the GSAT architecture. For further comparison of GALA, please use the updated results.

For DrugOOD datasets, we use the latest official release based on ChEMBL30 from https://drugood.github.io/ . For your information, in our previous work CIGA, we curated the datasets the ChEMBL29 following the DrugOOD repository.

Please feel free to let us know if you have any further questions!