GCNConv vs ChebConv

[chinmay5]

I am starting with the framework and would appreciate it if someone could help me with certain basic doubts:-

I have a node classification task and have used a ChebConv and GCNConv operators. While the ChebConv gives me decent results, (acc ~ 0.71), with the same code base and replacement by GCNConv, I am not able to converge the network. Here are some of my doubts and I would really appreciate it if someone could explain them:-

1. ChebConv would get neighborhood information based on the degree K. I should get the same effects by using K GCNConv layers cascaded. However, this does not seem to happen.
2. Since even the training loss is not decreasing much and I am using the same code base, there should be some specific property of ChebConv which sets it apart from GCNConv. Can someone please explain what that might possibly be?

[wsad1]

Its not easy to give a definitive answer here. The difference in accuracy might be down to your data or the number of layers you are using. Here are some thoughts on your 2 questions.

• ChebConv with degree K and GCNConv with K layers aren't the same. Each GCNConv layer is aggregating node embeddings from the previous GCNConv layer. But ChebConv always aggregates raw node embeddings.

• GCNConv tends to over-smooth representations when there are many layers. But ChebConv can potentially avoid that because of how its output is a "combination of k-hop node embeddings".

Two things you could try/check.

1. Look at the homophily of your labels. If the homophily is low, then that could be the reason why ChebConv works better.
2. Try adding JKconnections to your GCN that might help with the over-smoothing, if that is indeed what's causing the difference in accuracy.

[chinmay5]

Thank you, that helped me a lot. I found out that the homophily of the labels was ~0.5 and hence GCN struggled. Special thanks for the clarification with the first pointer. I replaced the GCN with SageConv and am currently experimenting with other suggestions but this undoubtedly gives me a better understanding of why ChebConv worked better. Thank you.