New benchmark tests 13 hyperbolic graph embedders for link prediction
Which hyperbolic embedding method finds missing links best? A unified benchmark reveals the winners.
Hyperbolic embeddings are a powerful way to represent complex networks in geometric spaces, but until now it's been hard to compare methods from different research communities. In this study, researchers benchmarked 13 unsupervised hyperbolic graph embedders under a unified protocol that tests both link prediction (recovering missing edges) and topology reconstruction (preserving local and global network structure). The evaluation spans synthetic and empirical networks, covering a wide range of structural regimes.
The results show that maximum-likelihood and representation-learning-based approaches, including hybrid variants, achieve the strongest overall performance. However, no single method dominates across all tasks and network types. Performance correlates more strongly with embedding paradigm than with whether the method came from machine learning, network science, or algorithmics. The paper identifies exactly where each paradigm succeeds or fails and gives practical guidance for choosing the right embedder in downstream applications like social network analysis, recommendation systems, and biological network modeling.
- Benchmarked 13 unsupervised hyperbolic graph embedders under a unified protocol for link prediction and topology reconstruction.
- Maximum-likelihood and representation-learning approaches (including hybrids) perform best, but no method wins everywhere.
- Embedding paradigm matters more than disciplinary origin; paper provides network-specific selection guidance.
Why It Matters
Gives data scientists a practical roadmap for choosing hyperbolic embedders in link prediction and network science tasks.