Hyperbolic Informed Embedding (HIE)

• HIE is a geometry-aware regularization framework that leverages cost-free hierarchical cues from hyperbolic distance to the origin.
• It combines root alignment and level-aware stretching to impose a hierarchy-consistent radial organization on embeddings.
• Empirical results show HIE can boost performance by up to 21.4% in tasks like link prediction and node classification on benchmark datasets.

Hyperbolic Informed Embedding (HIE) is a geometry-aware regularization framework for hyperbolic representation learning introduced in “Hyperbolic Representation Learning: Revisiting and Advancing” (Yang et al., 2023). It is designed to improve existing hyperbolic models by incorporating cost-free hierarchical information deduced from the hyperbolic distance to the origin (HDO), also called the induced hyperbolic norm. HIE is presented as both task-agnostic and model-agnostic, and its central mechanism is a combination of root alignment and level-aware stretching that encourages a more hierarchy-consistent radial organization of embeddings. The method was motivated by the claim that many prevalent hyperbolic models, despite being used on hierarchical or tree-like data, do not reliably place roots near the origin, do not produce the expected long-tailed radial distribution, and do not fully exploit the exponential capacity of hyperbolic space (Yang et al., 2023).

1. Origins, motivation, and scope

HIE emerged from a critique of a common assumption in hyperbolic representation learning: that hierarchy will be inferred automatically from pairwise similarity measurements or downstream task losses. The motivating paper argues that many prior hyperbolic models are trained with task-specific but geometry-irrelevant objectives, such as cross-entropy for node classification or link losses for link prediction, and that these objectives do not by themselves provide explicit information about roots, leaves, or hierarchical order (Yang et al., 2023). The resulting concern is not whether hyperbolic space can represent hierarchies, but whether ordinary optimization actually places the learned embeddings into a hierarchy-consistent configuration.

The paper supports that concern with a position-tracking mechanism based on HDO. On synthetic tree data, the authors report that the root is not near the minimum HDO: with HGCN, root HDOs were around 3.1 and 3.3, while the minimum HDOs were 2.0 and 2.1. They further report that the learned HDO distribution is “roughly normal rather than long-tailed,” and that the overall embeddings are “not maximally scattered” (Yang et al., 2023). These observations motivate HIE as an explicit training-time intervention.

The acronym is potentially ambiguous. A different paper titled “Joint embedding in Hierarchical distance and semantic representation learning for link prediction” introduces a knowledge-graph model also called HIE, but that model is not a hyperbolic embedding method and does not use the Poincaré ball, Lorentz model, or any explicit curvature parameter (Liu et al., 2023). In the hyperbolic-learning literature, HIE therefore refers specifically to the radial, origin-based regularization strategy of (Yang et al., 2023).

2. Geometric basis

The geometric premise of HIE is that, in a good hyperbolic embedding of hierarchical data, the root should be close to the origin, and nodes at deeper hierarchy levels should have larger hyperbolic norm. The method uses the hyperbolic distance to origin as a proxy for hierarchical level:

$d_{\mathcal H}(\mathbf{x}, \mathbf{o}).$

This quantity is called HDO and is also described as the induced hyperbolic norm (Yang et al., 2023).

The rationale is the standard tree–hyperbolic correspondence. In a $k$-regular tree, the number of nodes grows exponentially with level, and in hyperbolic space the area of a ball or disc also grows exponentially with radius. The paper therefore argues that “strategically placing the nodes of level $r$ of a tree at a distance $R$ (where $R$ is proportionate to $r$) from the hyperbolic origin results in a hierarchical embedding that captures the underlying tree-like structure” (Yang et al., 2023). HIE does not prove a formal equivalence between HDO and true hierarchy; it uses this geometric match as a training prior.

The framework is defined for both major constant-curvature models used in hyperbolic learning. For the Poincaré ball, the appendix gives

$$\mathcal{B}_\kappa^d=\{\mathbf{x}\in\mathbb{R}^d~|~ \|\mathbf{x}\|^2<-1/\kappa\},$$

with metric tensor

$$g_{\mathcal{B}^\kappa} = (\lambda_\mathbf{x}^\kappa)^2\mathbf{I}_d,\qquad \lambda_\mathbf{x}^\kappa=2(1+\kappa\|\mathbf{x}\|_2^2)^{-1},$$

and Möbius addition

$$\mathbf{x} \oplus_{\kappa} \mathbf{y} = \frac{ \left(1-2 \kappa\langle \mathbf{x}, \mathbf{y}\rangle_{2}-\kappa\|\mathbf{y}\|_{2}^{2}\right) \mathbf{x} + \left(1+\kappa\|\mathbf{x}\|_{2}^{2}\right) \mathbf{y} } { 1-2 \kappa\langle \mathbf{x}, \mathbf{y}\rangle_{2}+\kappa^{2}\|\mathbf{x}\|_{2}^{2}\|\mathbf{y}\|_{2}^{2} }.$$

For the Lorentz model, the paper gives

$$\mathcal{L}_\kappa^d = \{\mathbf{x}\in\mathbb{R}^{d+1}~|~\langle\mathbf{x},\mathbf{x}\rangle_\mathcal{L}= {1}/{\kappa}\},$$

with Minkowski metric tensor $k$0 (Yang et al., 2023).

This broader geometric context is consistent with other work that interprets radial position as a hierarchy signal. In text, hyperbolic norm was found to correlate with word-context frequency and phrase constituency (Dhingra et al., 2018). In image embeddings, points near the origin were interpreted as more generic or uncertain, and points near the boundary as more specific or confident (Khrulkov et al., 2019). These results suggest that HIE is part of a wider radial-hierarchy interpretation rather than an isolated regularizer.

3. Formal mechanism

HIE adds two operations to an existing hyperbolic learner: root alignment and level-aware stretching (Yang et al., 2023). It is therefore not a new manifold and not a new encoder architecture; it is an auxiliary mechanism attached to existing hyperbolic models.

The first component is the hyperbolic embedding center $k$1, used as a proxy root. The paper states that $k$2 is the solution of the weighted Fréchet-type center problem

$k$3

The corresponding tangent-space center is

$k$4

For the Poincaré ball, the paper gives a Möbius-gyromidpoint-style formula; for the Lorentz model it gives a weighted centroid formula (Yang et al., 2023).

Once the center is estimated, embeddings are aligned to the origin. In hyperbolic form,

$k$5

and in tangent space,

$k$6

After alignment, HDO is computed as a weighted average:

$k$7

where

$k$8

and the paper uses the identity function for $k$9 “for simplicity” (Yang et al., 2023).

The stretching loss is then

$r$0

with $r$1 monotone increasing. Minimizing this objective increases the average weighted HDO. The intended effect is asymmetric: high-level nodes near the origin receive small weights and are not pushed outward aggressively, whereas low-level nodes that are already far from the origin receive large weights and are stretched further into the high-capacity outer region (Yang et al., 2023).

The final training objective is

$r$2

with $r$3 in the reported experiments (Yang et al., 2023). The paper also distinguishes two implementation modes: a hard operation, which replaces the original embeddings with aligned ones, and a partial root-alignment, which applies alignment only inside the HIE branch to generate gradients. The reported experiments mainly use the partial version.

4. Empirical results and diagnostic effects

The reported experiments cover synthetic trees, shallow hyperbolic models, hyperbolic neural networks, and graph neural networks on Disease, Airport, Citeseer, and Cora (Yang et al., 2023). The headline result is that HIE achieves an improvement of up to 21.4% over competing baselines. That number comes from shallow hyperbolic link prediction on Disease with 25% training links and 256 dimensions, where the hyperbolic baseline AUC is reported as 55.0 ± 0.4 and HIE reaches 66.8 ± 0.4. In the same setting, AP improves by +20.4% (Yang et al., 2023).

A few representative results illustrate the reported scale of the gains:

Setting	Baseline	HIE
Disease link prediction, 25% links, 256 dim, AUC	55.0 ± 0.4	66.8 ± 0.4
Disease node classification, HNN++, 64 dim	67.4	78.4
Citeseer node classification, HGCN, 64 dim	67.6	74.1
Cora node classification, HGCN, 64 dim	78.5	83.0

The paper reports that HIE improves not only downstream scores but also the geometry of the learned embeddings. Mean HDO increases; more nodes move toward larger radii; the HDO distribution becomes more long-tailed; and the root or hyperbolic center becomes close to the minimum HDO (Yang et al., 2023). On synthetic trees, the paper also reports improvements in relative hierarchy accuracy: for TREE-L, from 75.3% to 77.6% at 256 dimensions, and for TREE-H, from 73.7% to 83.5% (Yang et al., 2023).

Ablation results show that the two components of HIE behave differently across graph types. On Disease, alignment alone gives a large gain and stretching alone can hurt. On Citeseer and Cora, stretching alone helps more than alignment alone. The full method is reported as best overall (Yang et al., 2023). This is consistent with the paper’s interpretation that tree-like data benefit especially from correct root placement, whereas less tree-like data benefit more from radial spreading.

The paper also includes an explicit failure case for the opposite operation. When nodes are pushed toward the origin rather than outward, performance collapses: Citeseer drops to 18.1 and Cora to 31.9 (Yang et al., 2023). This negative result is important because it shows that HIE is not merely a norm regularizer; it depends on the specific asymmetry between central roots and outer leaves.

5. Relation to the broader “informed” hyperbolic literature

HIE in the strict sense denotes the method of (Yang et al., 2023), but the broader research landscape contains multiple examples of hyperbolic embeddings that are informed by external structure, side information, or explicit hierarchy. These works are not identical to HIE, but they illustrate a common design pattern: hyperbolic geometry is strengthened when the model is given a structured signal beyond raw pairwise similarity.

In attributed networks, HEAT introduces a modified random walk with teleportation that mixes topology and node-attribute similarity, then learns hyperboloid embeddings from the resulting positive pairs (McDonald et al., 2019). In topic taxonomy mining, HyperMiner builds words and topics in hyperbolic space and adds a hyperbolic contrastive regularizer based on a WordNet-derived concept taxonomy (Xu et al., 2022). In hierarchical graph embedding, “A Geometry-Aware Algorithm to Learn Hierarchical Embeddings in Hyperbolic Space” identifies capacity, intra-subtree, and inter-subtree illnesses, and addresses them with dilation and transitive closure regularization (Wang et al., 2024). These methods suggest that HIE belongs to a broader class of geometry-aware strategies that inject hierarchy explicitly rather than hoping it emerges automatically.

Recent domain-specific systems extend the same pattern. HypRAG develops hyperbolic dense retrieval in the Lorentz model and reports up to 29% gains over Euclidean baselines on RAGBench, while also finding over 20% radial increase from general to specific concepts (Madhu et al., 8 Feb 2026). HEEGNet combines Euclidean and hyperbolic encoders with a domain adaptation module for EEG and is motivated by the claim that EEG exhibits hyperbolicity (Li et al., 6 Jan 2026). OpenHype uses a continuous hyperbolic latent space for hierarchical open-vocabulary radiance fields and performs hierarchy traversal by geodesics (Weijler et al., 24 Oct 2025). These systems do not implement HIE as defined in (Yang et al., 2023), but they reinforce the same principle: hyperbolic representations become more effective when hierarchy is made explicit in training, architecture, or supervision.

6. Misconceptions, limitations, and status

A common misconception is that HIE is a new hyperbolic neural architecture. It is not. The paper explicitly frames it as a lightweight add-on that “does not introduce additional model parameters or change the original architecture” (Yang et al., 2023). Its role is to regularize the geometry of already hyperbolic embeddings. A second misconception is that HIE proves hyperbolic norm is always a correct proxy for hierarchy. The paper does not provide a full formal theory establishing that HDO ordering always matches true hierarchy; its argument is geometric and empirical (Yang et al., 2023).

The method has several visible limitations. First, its benefit can be small when the base model already learns a good hierarchy. The paper identifies Airport as a case where improvement is slight because the baseline may already be “well hierarchically organized” (Yang et al., 2023). Second, the method assumes that a meaningful center or root proxy exists. On non-tree-like graphs, alignment alone can crowd points near the origin and hurt performance. Third, some implementation details are not fully specified in the main text; the paper notes, for example, that the exact choice of $r$4 is not fully reported across all experiments (Yang et al., 2023).

Finally, the term HIE should not be generalized indiscriminately. In one line of work it means the specific origin-distance regularizer of (Yang et al., 2023); in another, unrelated line it names a knowledge-graph model that is explicitly not hyperbolic (Liu et al., 2023). Within hyperbolic learning proper, HIE is best understood as a concise formulation of a larger thesis: hyperbolic space alone is not enough. If hierarchy matters, radial organization must often be made explicit through alignment, stretching, structured sampling, taxonomy-aware regularization, or other geometry-aware interventions (Yang et al., 2023).