search

yifan-h / CS-GNN

Measuring and Improving the Use of Graph Information in Graph Neural Networks
MIT License
82 stars 10 forks source link

The Experimental Results on PPI? #1

Closed mingloo closed 4 years ago

mingloo commented 4 years ago

Thanks for sharing the code! @yifan-h

Do you have some experimental results comparing CS-GNN vs GAT on PPI?

yifan-h commented 4 years ago

Hi Ming,

Sorry we don't evaluate CS-GNN on PPI dataset.

PPI dataset is composed of many small subgraphs. We cannot get the results for other models which don't support mini-batch. And our paper doesn't try to beat GAT with better attention mechanism. The main point is using two smoothness metrices to help researchers understand when and why GNNs work.

Besides, there are many interesting features of PPI dataset. The node features are fairly sparse, and we think they are generated by one-hot encoder. The attention coefficients are highly related to the cosine value of two connected nodes' features. There are many recent papers trying to figure out why GAT achieves significant performance on PPI. Hope you will find them helpful.

mingloo commented 4 years ago

Thanks for your prompt support.

Closed this issue now.