search

JHL-HUST / NAGphormer

NAGphormer: A Tokenized Graph Transformer for Node Classification in Large Graphs
Apache License 2.0
115 stars 15 forks source link

Unable to reproduce the results #3

Open kaansancak opened 1 year ago

kaansancak commented 1 year ago

Hi, I am unable to reproduce the results reported in the paper. I downloaded the datasets from GRAND+ and applied the preprocessing mentioned in the paper. See this. These are the results I get: 75.32 on Amazon2M, 0.4652 on Aminer, and 0.8965 on reddit. I followed the split instructions as well. Can you provide your splits for this data? Also the code provided here is not enough to reproduce the results. Your code assumes that the data is preprocessed and saved with its train/val/test masks. Although, there is no such split provided in GRAND+. Can you provide some clarifications?

guangnianchenai commented 1 year ago

Hi, please set renormalize=True when you download data from GRAND+.

kaansancak commented 1 year ago

Thanks! This resolves the problem. I opened a PR: https://github.com/JHL-HUST/NAGphormer/pull/4

zzyzeyuan commented 1 year ago

Thanks! This resolves the problem. I opened a PR: #4

Hi, I just downloaded this data from Google Drive and got nearly same results (75.32 on Amazon2M, 0.4652 on Aminer, and 0.8965 on reddit).

But I can't find where to set renormalize=True when I download datasets. Could you give me some hints?