Dear author,

I hit upon that it seems like the global pooling operation in graph classification is on the wrong axis. For example, when reading out the "out" ( graph_num, node_size, fea_num) with axis=-1, the resulting shape is (graph_num, node_size) but not the desired (graph_num, fea_num).

Hi

I attempt to run the code in "BA-shapes.ipynb" and got an error "ValueError: Input contains NaN, infinity or a value too large for dtype('float32')". Only "BA-shapes.ipynb" has this issue. The other three experiments are fine. I think this may be caused by the python environment. Could you please provide your environment for running experiments?

Here is my output of running "BA-shapes.ipynb" and "pip freeze": https://gist.github.com/haolanut/814d3908fd4a91fd37da22b0349269cc

Also, I'm wondering are you planning to provide the script for explaining the graph classification model?

Thank you!

Dear author, may I know when you will release the complete code for explaining GNN for graph classification tasks?

We have some difficulties in reproducing the experimental results in your paper. In particular, we have the following questions

Questions: 1: For the dataset MUTAG, do you use any node features when training the GNN model for graph classification tasks? And did you use the batch norm layer in the GNN model? If yes, what is the batch size?

2: When you calculate the AUC, how do you know the ground truth for explaining the GNN model for graph classification tasks? And when training the explainer, do you use the entire dataset, including both classes, or only use the mutagen graphs.

Hi,

Really cool project. I was wondering if you had plans to release the BA-2Motifs datasets as well. At the moment it is the missing from the datasets folder, and if you would be able to provide more details on the hyper-parameter choices you have made in the graph classification models.

While I looked at the appendix, I see that there are additional options in the model definition including (Batchnorm, Concatenating an add pool option etc).

I'm trying to replicate your results in Pytorch/ Pytorch Geometric for the graph classification set-ups.

Thanks!

the loss in the paper is: but in code: you just take out the probability of the item corresponding to the label,and use -tf.math.log() to get the pred_loss,$\sum_{c=1}^C$ I don't think it shows up in code.

Hi,

Really cool project. When I compared this implementation with the one decribed in the paper, I found out that this code does not sample K times for each node's subgraph. Is there any reason for it and will it leads to a high variance for a certain node?

Looking forward to your response!

Thanks!

Why the mean auc score calculated in BA-shapes.ipynb is about 99%. I also tried all nodes from three classes in house motif, the auc score is still around 99%. And that performs quite better than the 0.963±0.011 score you presented in the paper. Are the calculations of auc score different?

Hello. How are you? I'm investigating PGexplainer and I noticed that it is incredibly weak at explaining graphs with nodes with more than 1000 neighbors. An example is the bitcoin-alpha and -otc dataset. PGexplainer (with 3hop) does not achieve 1% explanation accuracy. For 1hop (average degree 17), PGexplainer achieves ~83% accuracy. Why does this occur?

Another point is the explanation time which is said to outperform GNNexplainer, however, for cases with nodes with degree greater than 1000, PGexplainer is much slower than GNNexplainer. Why does this occur?

Hello,

I was playing around with train_BA-2motif.ipynb when I realized that some input graphs are not as expected. Specifically, by running this code after the cell where you read the dataset:

for i in adjs:
    d = nx.from_numpy_array(i)
    if not (nx.is_connected(d)):
        nx.draw(d)
        plt.show()

I saw that there are about 40 graphs that are disconnected and in which the target motif has a self-loop. This behavior is unexpected after reading the dataset description of the original paper.