ponderly / pc-gnn Goto Github PK

I had some trouble reproducing GAT. Could the author open source the GAT code in the paper? Thank you very much~

Hi author, @PonderLY In your code, it used sparse matrix. When the graph is really huge, it cannot work. For a sparse matrix even will occupy more than 2TB memory.

How to solve that?

Thanks for your amazing paper!
But a problem bothered me when I read the code.
Compared with aggregation process in paper(eq.8 eq.9), weight matrix as well as l-1 layer r relation representations of central nodes seems to be missing in this code . For example,I can't find d_l *2d_{l-1} weight matrix in IntraAgg class.Also summation operation instead of concat operation is applied to inter_agg .
Could you help me with my problems?

Hi, l am trying to implement GCN on the Amazon dataset, but l have found a significant difference from your results. You said you adjusted the binary classification threshold on GCN,DR-GCN, and GraphSAGE. but l can't find it in the code. I don't know where to find your binaryclassification threshold strategy.

hello authors:

can you release M7 ~ M9 datasets by used in your paper, or teach us how to generate such datasets?

thanks!

Thank you for your exciting work!

I've got a question on the perforamance comparison in Table 3 after reading your paper:
For Amazon, the performance reported in Table 3 is (0.6416±0.0079, 0.7589±0.0046, 0.5949±0.0349) respectively.
But the result of the author implementation is inconsistent with Table 3:

Restore model from epoch 45
Model path: ./pytorch_models/2021-08-08 19-38-06/amazon_SAGE.pkl
   GNN F1-binary-1: 0.8481	F1-binary-0: 0.9848	F1-macro: 0.9165	G-Mean: 0.8963	AUC: 0.9324	Recall=0.8987
   GNN TP: 268	TN: 3110	FN: 62	FP: 34
F1-Macro: 0.9164508862362816
AUC: 0.9323704603284756
G-Mean: 0.8962916549983307

Additionally, I implement these models based on DGL, performance of GraphSAGE is:

[2021/08/08 19:52:01.888] Saving Model to /trained_models/fraud/amazon/SAGE_amazon.pth
[2021/08/08 19:52:01.955] AUC=0.8709 | Recall=0.8709 | Gmean=0.8646
[2021/08/08 19:52:01.955] TN= 3033 FP=   77 FN=   81 TP=  266
[2021/08/08 19:52:01.955] Best val AUC=0.8844, f1-macro=0.8844, epoch=47
[2021/08/08 19:52:01.955] Final scores: ACC=0.9543 | f1-fraud=0.7710 | f1-benign=0.9746 | f1-micro=0.9543 | f1-macro=0.8728

Both implementations use the same configuration, the number of neighbors sampled is set to be 5.

Is there any other setting or constrain we need to pay attention to ?

Any ideas on how to extend this to graphs with different types of nodes? Thanks!

Hello, author. I noticed that you used oversampling when sampling fraud nodes in the training phase, requiring you to select nodes that are consistent with the fraud node label. This procedure requires the type label of the node. However, it is not possible to obtain the node type label in the inference stage, so it is impossible to oversample the fraudulent nodes, which leads to the inconsistency between the sampling method and the training stage, and the model in the training stage is more ideal. Does this not fit the logic of machine learning? Causing the trained model to be unreliable?

Hi guys,
I find an Error when calculating the label frequency of class C(v)
The code is in the src/util.py line 111
the original code is:
lf_train = (y_train.sum()-len(y_train))*y_train + len(y_train)
but i think it should be:
lf_train = (2 * y_train.sum()-len(y_train))*y_train + len(y_train) - y_train.sum()

is that right?

thank you!