City data combines sources with very different shapes. Points of interest (POIs) sit in category trees: food, then restaurant, then ramen shop. Road networks branch. Mobility flows follow power laws, so a few hubs dominate many minor nodes. Traffic speeds and travel times, by contrast, are plain continuous numbers. Hyperbolic space is a curved geometry whose volume grows exponentially with radius, so it behaves like a continuous tree, and it can store hierarchies in a handful of dimensions where flat space needs hundreds. Urban models that do use it currently either rebuild the whole network inside curved space, or apply it to one component and then drop back to ordinary flat vectors for sequence modelling. Nobody has asked which of these is necessary.
This note proposes a backbone-agnostic hyperbolic interface. Hierarchical urban sources enter and leave the model through hyperbolic embeddings, and an ordinary flat backbone sits in between: a Transformer, a state space model such as Mamba, or a graph network. A consistency loss forces the backbone's outputs to keep the input hierarchy intact. The central claim is testable: if hierarchy is enforced at the model's boundary, most of the benefit of fully hyperbolic architectures should carry over to any backbone family. If that claim fails, we learn something equally useful, namely that curved geometry must be native to the sequence operator itself.
Core Claim
Hyperbolic models for heterogeneous urban data treat geometry and architecture as one package. HMamba rewrites state-space equations in hyperbolic form, and Hypformer and THG rewrite attention. Each is a single bespoke architecture for one kind of sequence. At the other extreme, HVGAE encodes the static POI graph with a hyperbolic graph network and then models check-in sequences with a Euclidean Mamba. In that design hierarchy enters once and is never checked again.
Our claim sits between these two designs. It has two parts. First, the hierarchical part of the signal can be carried through an arbitrary flat backbone, provided that the backbone's outputs, mapped back onto the hyperbolic manifold, are required to preserve the input's radial ordering (general items near the origin, specific ones near the edge) and its tree distances. Second, because this constraint lives only at the interface, its benefit should transfer across backbone families with very different internal biases.
That second part is what ties the work to architectural generalization. Methods built on one backbone often fail on another. For example, prototype regularizers that work on CNNs break on Vision Transformers, as reported in Prototype-based Consistency Regularization. Our hypothesis is that a geometric constraint on what goes into and comes out of a backbone is less sensitive to that backbone's internal structure than a constraint on its intermediate features.
Research Program
The interface has three parts.
The first is an encoder for each source. The POI category tree and the road hierarchy are embedded in the Lorentz model, a numerically stable form of hyperbolic space that avoids the boundary blow-up of the Poincaré ball (Lorentz model). These sources use a learnable curvature per source, following the curvature-sharpness link established in Curvature Learning for Generalization of Hyperbolic Neural Networks. Continuous signals such as speed, time, and coordinates stay Euclidean. This split is deliberate. We do not force everything into one curved space, because continuous quantities have no tree structure to exploit.
The second part is a map into the backbone. Hyperbolic tokens are passed through the logarithmic map at the origin, which unrolls the curved space into an ordinary vector, then concatenated with the Euclidean tokens and fed to any backbone unchanged.
The third part is a hyperbolic readout. Backbone outputs are mapped back to the manifold with the exponential map, and two losses are applied there. The first is the task loss. The second is a hierarchy-consistency loss:
$$\mathcal{L}_{\text{hier}} = \sum_{(u,v)} \big(d_{\mathbb{H}}(z_u, z_v) - \lambda\, d_T(u,v)\big)^2 + \beta \sum_{u \prec v} \max\big(0,\ \|z_u\|_{\mathbb{H}} - \|z_v\|_{\mathbb{H}} + m\big).$$
Here $z_u$ is the readout embedding of entity $u$. The term $d_{\mathbb{H}}$ is hyperbolic geodesic distance and $d_T$ is distance in the category tree or road hierarchy. The scalar $\lambda$ is a learned scale. The relation $u \prec v$ means $u$ is an ancestor of $v$. The quantity $\|\cdot\|_{\mathbb{H}}$ is distance from the origin. The margin $m$ and weight $\beta$ are hyperparameters.
The first term preserves tree distances. The second keeps general concepts closer to the origin than specific ones. This mirrors the radial-hierarchy behaviour that HMCL showed collapses when it is left unconstrained.
The analysis has two strands. The first is a placement grid. We cross four geometry placements with four backbone families. The placements are fully Euclidean, hyperbolic input only (HVGAE-style), our interface, and a native hyperbolic backbone. The backbones are a GPT-2-style Transformer as in BIGCity, Mamba, a relation-aware graph convolutional network as in HREP, and the linear graph attention of STBP. Native hyperbolic versions exist only for the attention and state-space families, via Hypformer and HMamba. The second strand is probing. At each backbone layer we measure how well hierarchy survives inside the flat hidden states, by comparing tree distances against distances recovered after applying the exponential map. This tells us whether Euclidean backbones scramble hierarchy internally and the readout repairs it, or whether hierarchy survives unaided.
Baselines
The naive reference is a fully Euclidean model at matched parameter count on each backbone. Every source is embedded in flat space and fused by concatenation followed by the backbone. Any method that cannot beat this has no case.
The strongest existing methods differ by benchmark.
On next-POI recommendation the strongest competitor is HVGAE. It is the closest design: a hyperbolic POI encoder feeding a Euclidean Mamba. It reports HR@10 of 0.8421 and NDCG@10 of 0.7580 on Foursquare NYC. We will rerun it with the authors' preprocessing, which filters POIs with fewer than 5 check-ins. Beating it means exceeding both numbers on NYC and also winning on TKY and Yelp under our own rerun, since the file gives no numeric values for those two datasets.
On the Manhattan region benchmark the strongest competitors are HREP and MGFN. The file gives only relative improvements for them: HREP improves crime-prediction MAE by 5.23% over its best baseline, and MGFN improves check-in R² by 13.11% over MVURE. We found no reported absolute result on Lacuna for either method on Manhattan, so both will be rerun under our protocol.
On trajectory and traffic tasks the competitor is BIGCity. The file reports only relative gains, about 2.83% over GWNET and ST-Norm, so we again found no reported absolute result on Lacuna.
Native hyperbolic backbones, HMamba and Hypformer, are the upper reference for the placement question. We do not need to beat them. The interface needs to come within a stated margin of them while also being usable on backbones that have no hyperbolic version.
The key ablation is a Euclidean twin of our method. It uses the same readout losses, but with Euclidean distance and Euclidean norm in place of their hyperbolic counterparts, at identical dimension. If this twin matches our method, the gain comes from supervising hierarchy, not from curvature.
Fairness rules apply to every comparison:
- All methods use the same data splits and matched parameter budgets.
- Embedding dimensions are swept over 8, 32, 128, and 512.
- Each method gets an equal hyperparameter search of 30 trials, with 5 seeds per configuration.
- Results are reported as mean and spread.
Why not just…?
Why not just make the Euclidean embeddings bigger? Low-dimensional comparisons favour hyperbolic space strongly. HHNE reports about 0.87 link-prediction AUC in two dimensions on DBLP, against about 0.67 for the best Euclidean baseline. Euclidean models there needed 128 or more dimensions to catch up. Even so, a large Euclidean model may simply close the gap at the sizes people actually deploy. Our dimension sweep tests this directly. If the Euclidean baseline matches us at 128 or 512 dimensions, we will report that the interface's value is limited to compression and low-dimensional regimes.
Why not just add a hierarchy loss in flat space? This is the most serious threat. Much of the benefit could come from telling the model about the category tree at all, regardless of geometry. The Euclidean-twin ablation is built to settle it. The volume argument from Poincaré Embeddings predicts that the flat twin will fail specifically at deep, high-branching leaves, where crowding occurs. We will therefore report error stratified by POI depth, not only aggregate metrics. If the twin matches us at every depth, the paper's conclusion becomes "supervise hierarchy" rather than "use hyperbolic geometry."
Why not just use a hyperbolic encoder and stop there, as HVGAE does? HVGAE's own ablation shows that hyperbolic encoding helps. However, nothing in that design prevents the downstream Euclidean sequence model from scrambling the hierarchy. HMCL's observation of radial drift in curved models suggests such scrambling is plausible. The "hyperbolic input only" row of the grid, combined with the layer-wise probe, tests whether the readout constraint adds anything beyond the input encoding.
Why not just build fully hyperbolic backbones? Native designs exist only for particular operators. Each new backbone family would need its own derivation, which is exactly the portability problem this note targets. If native models clearly beat the interface wherever both exist, that is an informative negative result, and we will state it as such.
Proposed Experiments
We will use three benchmarks:
- Foursquare NYC and TKY plus Yelp for next-POI recommendation, scored by HR@K and NDCG@K for K of 5 and 10. The Foursquare category trees provide the POI hierarchy.
- The Manhattan dataset, with 180 regions, 20,000 POIs, and about 10 million taxi trips. We report MAE, RMSE, and R² on crime and check-in prediction, and NMI and ARI for land-use clustering.
- The BIGCity Beijing, Xi'an, and Chengdu data for next-hop prediction, travel time estimation, and traffic forecasting. We also run a Beijing-to-Xi'an transfer test of whether hierarchy-preserving readouts generalize across cities.
Three outcomes would be decisive.
First, the interface beats the Euclidean twin on deep-leaf POIs, and on at least three of the four backbones.
Second, the interface comes within 2% relative of HMamba and Hypformer on the two families where they exist. We set this margin in advance.
Third, the gain is larger on hierarchical tasks (next-POI, land use) than on continuous ones (traffic speed). We expect this because the geometry only targets tree-shaped structure. If gains were uniform across task types, that would suggest the method works for some other reason, and we would investigate it.
We will separately report the layer-wise hierarchy-distortion probe for each backbone. It is the most likely place for a genuine architectural finding: which backbone families keep hierarchy intact on their own, and which destroy it.
References
- HMamba: Hyperbolic Mamba for Sequential Recommendation
- Hypformer: Exploring Efficient Hyperbolic Transformer Fully in Hyperbolic Space
- THG: Transformer with Hyperbolic Geometry
- Hyperbolic Variational Graph Auto-Encoder for Next POI Recommendation
- Semi-supervised Semantic Segmentation with Prototype-based Consistency Regularization
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Curvature Learning for Generalization of Hyperbolic Neural Networks
- Hyperbolic Multimodal Continual Learning
- BIGCity: A Universal Spatiotemporal Model for Unified Trajectory and Traffic State Data Analysis
- Heterogeneous Region Embedding with Prompt Learning
- A General Spatio-Temporal Backbone with Scalable Contextual Pattern Bank for Urban Continual Forecasting
- Multi-Graph Fusion Networks for Urban Region Embedding
- Hyperbolic Heterogeneous Information Network Embedding
- Poincaré Embeddings for Learning Hierarchical Representations