Frequency Adaptive GCN (FAGCN).

[wsad1]

🚀 Feature

FAGCN was proposed in this paper.

Motivation

According to the paper FAGCN works especially well on graphs with low homophily(disassortative). This is due to the fact that it can adaptively use low and high frequency signals in the graph.

Additional context

The implementation looks straightforward. But i wanted to get peoples thoughts before creating a PR.

[rusty1s]

Very much appreciate a PR on this one :) Let me know if you need any help.

[wsad1]

Thanks @rusty1s . So i'd like some thoughts on the implementation.
Specifically. This is how the full FAGCN layer is implemented.

The raw features are transformed , using an MLP, and added to the aggregation at every layer. Now forward could be implemented in 2 ways.

1. forward(x,h0). Here a user would have to pass node embedding from the previous layer and the transformed raw features. And forward would perform all the operations shown in the image, except for the MLP part.
2. forward(x). Here forward would only perform the aggregation operation and it would be left to the user to implement the rest of it, they would have to add transformed raw features to forwards output, then they would have to do the last Linear transform too.

Which do you think is better.

[rusty1s]

I would go with the first option, similar to what we do in GCN2Conv.

[wsad1]

Great, thanks.