Octavian-Eugen Ganea, Gary Bécigneul and Thomas Hofmann (2018), Proceedings of the 35th International Conference on Machine Learning (ICML 2018), PMLR 80:1646–1655 — ARXIV-1804.01882
Key takeaways
- A partial order as nested regions. Each concept is a point x of the Poincaré ball Dⁿ. Its entailment cone S_x is the exponential map of the tangent vectors at x within angle ψ(x) of the outward ray from the origin through x. The relation “v is in S_u” is the model’s “u entails v” (v is a subconcept of u). If cones nest, so that v in S_u implies S_v ⊆ S_u, the relation is transitive. Points in a cone are farther from the origin than its apex, which makes it antisymmetric. That second step is my derivation; the paper states only that the cones “induce a partial ordering”. It generalises Vendrov et al.’s order embeddings, where x is more general than every y with yᵢ ≥ xᵢ for all i.
- What nesting forces (§3, proofs in Appendices E–F). Ask for cones that are symmetric about the outward ray, have an aperture depending only on the apex’s radius, and vary continuously. Then transitivity forces ψ ≤ π/2: a cone covers at most half the ball (Lemma 2). It also forces h(r) = r sin ψ(r)/(1 − r²) to be non-increasing (Theorem 3). Since h → 0 as r → 0, no non-trivial cone can be defined near the origin. The cone family is therefore restricted to ‖x‖ ≥ ε, with no top element. Keeping h at its largest value K gives the widest apertures, sin ψ(x) = K(1 − ‖x‖²)/‖x‖, with K ≤ 2ε/(1 − ε²) (Eqs. 24–26). Cones narrow toward the boundary, where specific concepts sit.
- The closed form. For this ψ, transitivity holds (Theorem 4), but the proof is given only as “similar to that of Thm. 3”, and there is no appendix for it. Membership is the angle test Ξ(x, y) = π − ∠Oxy ≤ ψ(x), with Ξ in closed form (Eq. 28, Theorem 5), which I checked numerically against the hyperbolic law of cosines. The Euclidean analogue has sin ψ = K/‖x‖ and its own Ξ (Eq. 31). The paper also derives closed-form unit-speed geodesics and the exponential map of the Poincaré ball (Theorem 1, Corollary 1.1). The experiments do not use that exact map, because it brought no significant gain over Nickel and Kiela’s retraction.
- Several parents are representable; a lattice is not. Cones of incomparable points can overlap, so a point in both lies below two apexes that neither entails the other. A DAG with shared descendants is therefore expressible, and the method is stated for any DAG. But the intersection of two cones is not in general a cone, so the order on the ball has no meets or joins. Nothing is proved about which posets embed, in what dimension, or with what margin. All of this is my analysis; the paper does not discuss it.
- The evidence is one link-prediction experiment (§5, Table 1). The data is the transitive closure of the WordNet noun hierarchy, with the root removed: 82,114 nodes and 661,127 edges. The training set always holds the transitive reduction, plus 0%, 10%, 25% or 50% of the other closure edges. Test F1 is on held-out closure edges against ten corrupted negatives each, at dimensions 5 and 10. Hyperbolic cones score 80.1–92.8 (dimension 5) and 85.9–94.4 (dimension 10) from 10% up, ahead of every baseline. At 0% every method scores 27–43, and order embeddings lead. So a transitive geometry trained on the reduction alone did not reconstruct the closure. Euclidean cones initialised from Poincaré embeddings perform “on par with” hyperbolic cones (stated without numbers). All results are single runs with no variance, and there is no full-reconstruction experiment.
Standing in the record
Filed on 2026-10-09 at the owner’s request. The owner added it to the batch of six works on hierarchy and hyperbolic geometry, asked for right after the record opened QUESTION-025, once the other six had been read. That question asks whether attributes that imply or exclude each other in co-occurrence still give linear attribute directions, and a concept lattice that is not Boolean. It builds on THEORY-185 and LIT-863. The six are Cagnetta et al.’s random hierarchy model (LIT-877), Krioukov et al. 2010 (LIT-865), Sala et al. 2018 (LIT-874), Lin et al. 2023 (LIT-870), Zhang et al. 2023 (LIT-876) and Yang et al. 2023 (LIT-871). It was read on its own merits (NOTE-671), and the reading produces THEORY-189.
Of the batch, it comes nearest the Lattice Representation Hypothesis (LIT-267), whose picture is implication as containment of regions. Sala et al. represent a tree by distances. This paper represents a partial order with multiple parents by asymmetric containment of cones. It does not answer QUESTION-025. It has no co-occurrence statistics and no attributes: each concept is a free point trained on WordNet hypernym edges. Its cones are not half-spaces, and no direction stands for an attribute. Its Euclidean baseline, order embeddings, is the half-space picture: concepts are conjunctions of coordinate thresholds. That baseline is the one the cones improve on once closure edges are in training. For CLAIM-119, it shows only that a non-tree partial order has a natural home in hyperbolic space. Whether a concept lattice embeds there, with its meets, is not addressed.