Open lettlini opened 1 month ago
I think this is a great feature request :)
Great! You are welcome to assign this problem to me and I will try to work on it in the next 2 weeks.
I now have a preliminary implementation of both the Dirichlet Energy and the Mean Average Distance. The code can be found here in my node_similarity_pyg
repository (I plan to add a notebook on how to replicate some of the empirical results of the paper in a few days).
I don't really know where in pytorch-geometric
these methods would fit best, maybe in torch_geometric.utils
?
Nice :) Yes, they would fit best in torch_geometric.utils
.
🚀 The feature, motivation and pitch
I recently came across the paper A Survey on Oversmoothing in Graph Neural Networks and thougt that having a ready-to-use implementation of the Dirichlet energy and the Mean Average Distance would make diagnosing oversmoothing much easier.
Alternatives
I saw that
PairNorm
has been implemented but I think it'd be nice to be able to quantitatively examine the behavior of node similarity (at least for academic & research purposes).Additional context
If this is something people would find useful if it was included in
pytorch_geometric
I could work on this.Edit: Corrected paper link