Questions regarding the graph classification tutorial on dgl website? (GCN vs GIN)

Questions
#1

So i was reading the latest graph classification guide :
https://docs.dgl.ai/en/latest/guide/training-graph.html

and it is really confusing me, first off its making a classifier that is using graphConv, which is GCN. then it loads a GIN dataset with GinDataset(), which got me confused. why are we using a GIN dataset with a GCN classifier, and why do we even need to have a GIN specific dataset? doesn’t GIN work with any type of graph dataset for the classification task?

then the tutorial goes on to say that " DGL implements GIN as an example of graph classification", but what does this even mean? the classifier that was built in this tutorial is using graphConv, what does GIN has to do with this? this is really confusing me so can someone clarify these stuff for me? sorry for the rookie question I’m new to all the graph neural network stuff :frowning:

#2

Sorry for the confusion. The GINDataset is nothing but an interface for datasets considered in How Powerful are Graph Neural Networks? along with the work-specific dataset settings. We should definitely clarify about that in the doc. I’ve opened a PR to address it and feel free to provide more suggestions.

For “DGL implements GIN as an example of graph classification”, will it make better sense to say “For a complete end-to-end example on graph classification, see GIN.”? Copy @BarclayII.

#3

Thank you for clarification, the only other suggestion i can make is please make more tutorials on graph classifications, for example showing how can we use GIN or other algorithms to do classification instead of just graphConv (or just simply updating the same tutorial and adding just few lines to show how the same thing can be done with other algorithms), so rookies like me can learn faster. thanks!

1 Like