bupt-gamma / graph-structure-estimation-neural-networks Goto Github PK

Hi! I am working on reproducing the results of Geom-GCN under the setting where there are 20 nodes for each class, just like yours.
For citeseer dataset, under the 0.6-0.2-0.2 setting in their work, my results are exactly the same as theirs. But under the limited data scenario in your paper(20 nodes each class), using the original hyper-parameters leads to acc 0.5, 15 percent lower than your report(0.65).
Did you meet the same situation when doing experiments on Geom-GCN? Is this due to hyper-parameters? I would be grateful if you shared with me the hyper-parameters settings you took for Geom-GCN.

Hello!

I have a question about Table 2 in your paper that all node classification accuracies are reported as acc (± sigma).

My question is that as "± sigma" is generated from different "torch manual seed", do we need to split the dataset differently according to different "seed"? Or in another word, when experimenting on certain method and dataset multiple times, does the dataset share the same train/val/test split?

Thank you very much!

hi, thanks for sharing this code. I am reading the corresponding paper and confused of eq (6). how to get this eq, and can you provide a more detail process?