[Feature Proposal] Attract-Repel Embeddings for PyTorch Geometric

[tommyly201]

[Feature Proposal] Attract-Repel Embeddings for PyTorch Geometric

Hi PyG team,

I'm interested in contributing an implementation of Attract-Repel embeddings from the paper "Pseudo-Euclidean Attract-Repel Embeddings for Undirected Graphs" (Peysakhovich et al.). https://arxiv.org/pdf/2106.09671

This approach addresses a fundamental limitation in traditional graph embeddings: their inability to effectively represent non-transitive relationships. By splitting node representations into "attract" and "repel" components, the method improves link prediction performance, especially on heterophilic graphs.

The implementation would include:

1. Core AR embedding layers that can be plugged into existing models
2. AR versions of common layers like GCNConv
3. Utilities for calculating R-fraction and other metrics
4. Example notebooks demonstrating improvements on link prediction tasks

The method has shown 10-20% AUC improvements on heterophilic networks while requiring minimal architectural changes. Result of a rudimentary implementation here: https://substack.com/home/post/p-157861370

Would the PyG team be interested in such a contribution? I'd appreciate any feedback or guidance before submitting a PR.

Thanks,
Tommy

[akihironitta]

Duplicate of #10084.