Hi authors, I have read your paper, which is quite interesting. Thank you for your great work.

But I have a question about the split of Cora Dataset.

I count the node number of train_mask, val_mask, test_mask in https://github.com/chennnM/GCNII/blob/ca91f5686c4cd09cc1c6f98431a5d5b7e36acc92/process.py#L157 which are 1192, 796, 497. The sum of nodes [train_mask, val_mask, test_mask] is not 2,708, which is different from nodes shown in your paper.

You can reproduce this phenomenon by the code: print('train_mask is %s' %train_mask.numpy().sum()) print('val_mask is %s' %val_mask.numpy().sum()) print('test_mask is %s' % test_mask.numpy().sum())

I don't understand why this happen. Could you please point out? Hope for your response. Thanks!

Hi @chennnM

Thanks for sharing the GCNII code.

I've found there is some accuracy gap between PyG implementation and pure PyTorch implementation on Cora dataset. See below table for details.

| Cora | Layers | Metric | |---------|--------|--------| | PyTorch | 64 | 88.49 | | PyG | 64 | 85.00 |

Is there any possible parameter setting/code difference might cause the accuracy gap?

Great work！We are reproducing the statistical results in Table 2 and want to know how to design the 100 runs mentioned in the article for GCN and GCNII. Use random seeds and fixed hyperparameters in Table 6? Or some other method to select 100 runs?

The results seems to be very good, cons. We have re-run the code and get the similar results with fixing the seed=42. We wish to cite this work with reporting the given results on these three standard datasets. Before that, there are some questions:

1. The reported results in Table 2 are just with the fixed seed? or with random seeds and run many times?
2. If you choose the best model (with hyper-parameters) based on the validation set in the experiments in Table 2, what's the setting of Table 3, like fixing the seed or others?

thanks,

Hey,

I found the OBG implementation only has ``initial connection'' but without identity mapping.

Furthermore, in the paper, you propose to use an APPNP like propagation method, i.e.,

(1-alpha)LH^{(\ell-1)} + \alpha X,

but in the implementation you are using

LH^{(\ell-1)} + alpha X.

Do I miss something? Or there is a specific reason you do this?

hello, when I python full-supervised.py, the error message occurs: cuda:0 pretrained/673e5d3d5b814bfca489d053f0f3a92f.pt Traceback (most recent call last): File "full-supervised.py", line 117, in acc_list.append(train(datastr,splitstr)) File "full-supervised.py", line 73, in train features = features.to(device) RuntimeError: CUDA error: out of memory

my cuda10.0 torch1.3.1+cu100,python3.6, do you know how to solve it?

Hi authors, I have read your paper recently.But I have some questions about the conclusions you get in the paper.How to get the conclusion that the convergence rate of node is related to degree from theorem 1?Except for the stability vector π,the part of the formula that represents the convergence process seems to have nothing to do with degree.Could you please point out? Thanks!

Hi! I have reproduced your paper recently, it's an interesting paper. But I get an 83.6% accuracy for Cora with your code, the parameters are the same as you provided. I am curious about the reason.

I'm looking forward to your reply! Thank you!

The paper states that Lemma 1 is taken from (Chung, 2007). However, the closest thing I found in (Chung, 2007) is Lemma 3 ( inequalities (51), (52) ), in which the upper bound contains (beta_k)^2/8 not (lambda_G)^2/2. I tried to derives Lemma 1 in the paper from Lemma 3 in (Chung, 2007) but have not succeeded.

I would really appreciate if you can provide a proof or point out any sources that lead to the proof of Lemma 1 in the paper. Thank you!.

(Chung, 2007): Four proofs for the cheeger inequality and graph partition algorithms, ICCM 2007

Hi, I want to test the performance of adding identity mapping. I try to comment out output = theta*torch.mm(support, self.weight)+(1-theta)*r, and add the line output = support. Now it's APPNP, right? But the result is the same as before 85.7% , why?