Hyperbolic Graph Representation Learning: Embed in One Metric, Optimize with Another
The paper separates where graph embeddings live from the metric used to train them.
The authors argue that standard hyperbolic optimization can stall at large radii because angular updates become too small. They frame alternative update rules as a one-parameter family of preconditioners, while keeping the embedding space at curvature -1. In their tree experiments, a two-stage scheme beats the best single-curvature choice, cutting loss by 46-74%. ArXiv · AI/CL/LG's note
The authors argue that standard hyperbolic optimization can stall at large radii because angular updates become too small. They frame alternative update rules as a one-parameter family of preconditioners, while keeping the embedding space at curvature -1. In their tree experiments, a two-stage scheme beats the best single-curvature choice, cutting loss by 46-74%. ArXiv · AI/CL/LG's note
score 4