search

gospodima / Extended-SimGNN

A PyTorch Geometric implementation of SimGNN with some extensions.
MIT License
57 stars 14 forks source link

About "Fully Connected Layers" Settings #11

Closed Sangs3112 closed 1 year ago

Sangs3112 commented 1 year ago

Hi, I noticed that in the SimGNN paper, there was a setting for FCL, "use 4 fully connected layers to reduce the dimension of the concatenated results from the NTN module, from 32 to 16, 16 to 8, 8 to 4, and 4 to 1." However, I found that in this code, there are only two fully connected layers, which are 32 --> 16, and 16--> 1. When I modify the code according to the Settings in the paper, I find that the performance is worse by ten times, I don't know why, isn't the general Settings in the paper better? Here is the code in this Repository lu1 lu2

Here is my code my1 my2

gospodima commented 1 year ago

Hi @Sangs3112,

If I correctly remember, my tests with more FCLs didn't bring markable improvements in results and I kept only 2 layers.

Sangs3112 commented 1 year ago

Thank you for your reply. Yes, you are right, and I am wondering why the performance was reduced according to the Settings of the paper. In addition, I found that even if I kept only two FCLs and used GIN with better performance than GCN, the results were still not comparable to the results in the paper.

gospodima commented 1 year ago

There were several issues regarding configuration, achieved performance and how they are comparable to results in paper: #3, #5. For example, pairwise node comparison caused significant overfitting effect for me. You can take a look on them.