nucleation

LIT-871: Hyperbolic Representation Learning: Revisiting and Advancing

3 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-669); worth reading as an empirical check, on synthetic trees and four benchmark graphs, of the assumption that a hyperbolic model trained on an ordinary task loss lays out the data's hierarchy by itself, root near the origin and leaves near the boundary. It does so only partly. On two 3-ary trees of 1,093 nodes, a hyperbolic graph convolutional network (HGCN) trained for link prediction or node classification puts the root nowhere near the smallest distance to the origin, and the nodes' distances to the origin come out roughly bell-shaped rather than piling up at the leaves; read off as levels, those distances order random node pairs correctly 69–75% of the time. The proposed fix, HIE, recentres the embedding on its hyperbolic centroid and adds a loss that pushes every node outward in proportion to its current distance. It raises that ordering to 75–84% and lifts benchmark scores, most on the least tree-like graphs (Citeseer, Cora). There is no theory of why task losses fail to produce the hierarchy, and the fix itself uses no level information.
Tagsrepresentation-learning, network-science, anthology-candidate
Cited as “paper” byNOTE-669 — Hyperbolic Representation Learning: Revisiting and Advancing
Cited as “source” byTHEORY-421 — In HGCN embeddings of two synthetic 3-ary trees trained for link prediction or node classification, distance from the origin is only a partial reading of hierarchical level: the root is not innermost (distance 1.8–3.3 against a minimum of 1.2–2.4) and level order is recovered for 69–75% of node pairs; a loss built only on pairwise distances leaves the origin arbitrary, THEORY-825 — Hyperbolic graph embedding models trained on task losses that never mention the hierarchy do not place a tree's root nearest the origin, and their distance from the origin orders hierarchical levels only partly

Graph View

Menglin Yang, Min Zhou, Rex Ying, Yankai Chen and Irwin King (2023), Proceedings of the 40th International Conference on Machine Learning (ICML 2023), PMLR 202:39639–39659 — ARXIV-2306.09118

Key takeaways

  • The assumption tested. Hyperbolic embedding methods are trained on losses that never mention the hierarchy: a softmax over hyperbolic distances for shallow embeddings, cross-entropy for node classification, a Fermi–Dirac link likelihood for link prediction. The expected outcome, root near the origin and each level further out, is assumed to arise by itself. The paper measures it with one statistic, the hyperbolic distance of each node to the origin (HDO), read as its level.
  • What it finds. On two synthetic 3-ary trees of eight levels (1,093 nodes) that differ only in how labels follow the tree, HGCN trained for link prediction or node classification leaves the root at HDO 1.8–3.3 while the minimum is 1.2–2.4. The HDO distribution is roughly normal, where a tree’s leaf-heavy level counts would put most nodes far out. Pairs of nodes ordered by HDO match their true levels in 69–75% of 5,000 random pairs (Table 9). The geometry has room for the tree, but the trained embedding uses the room only partly.
  • The fix, and what it does. HIE recentres the embedding on its hyperbolic centroid (the Möbius gyromidpoint or Lorentzian centroid), taken as a stand-in for the root, and adds a loss that rewards a large HDO-weighted mean of HDO, pushing every node outward in proportion to how far out it already is. No level labels enter. It raises the pairwise level ordering to 75–84%, and it improves link prediction on DISEASE by up to 21.4% (with 25% of links for training) and node classification on four graphs. The largest classification gains are on Citeseer and Cora, the least tree-like, which the authors put down to the outward push separating classes rather than to hierarchy.
  • What it does not give. No account of why task losses fail to produce the radial ordering, and no measure of hierarchy on the real graphs, whose levels are unknown. The evidence that HIE recovers hierarchy rather than just spreading the embedding is Table 9 alone, on the two synthetic trees.

Standing in the record

Filed on 2026-10-09 at the owner’s request, in a batch of six works on hierarchy and hyperbolic geometry asked for right after the record opened QUESTION-025: whether correlated or hierarchical attributes in co-occurrence still give linear attribute directions, and a concept lattice that is not Boolean. That question builds on THEORY-185 and LIT-863. The batch’s other works are Cagnetta et al.’s random hierarchy model, Krioukov et al. (2010) on the hyperbolic geometry of complex networks, Sala et al. (2018) on representation trade-offs for hyperbolic embeddings, Lin et al. (2023) on multiscale geometry recovering a latent hierarchy, and Zhang et al. (2023) on hyperbolic geometry in hippocampal representations. This paper cites Krioukov et al. for the tree–hyperbolic-space analogy that motivates its statistic, and Sala et al. among works on hyperbolic embedding.

Read on 2026-10-09 (NOTE-669). It does not answer QUESTION-025. It encodes hierarchy as distance from an origin rather than as linear directions, and it trains graph models rather than embedding co-occurrence. What it adds is a caution that applies to any proposed answer. A representation space that can hold a hierarchy, here a geometry built for trees, does not show that a trained representation does hold it. The hierarchy has to be measured directly, as the measurement arm of the question asks.

Backlinks