Does Curvature Itself Help? A Same-Code Audit of Hyperbolic Graph Networks

Hyperbolic embeddings place data in a curved space where the room available grows exponentially as you move away from the center, much as the number of nodes in a tree grows with depth. Graph networks built in this space report large wins on hierarchical graphs. Yet every audit of a reported "hyperbolic beats Euclidean" result has found the same kind of problem. The best-known example comes from WordNet: a 200-dimensional Euclidean baseline scored 0.168 mean average precision, against 0.823 for 5-dimensional Poincaré embeddings. That gap came from an accidental unit-norm cap on the Euclidean model and from an untuned learning rate. With both fixed, the Euclidean model reaches 0.922 (Comparing Euclidean and Hyperbolic Embeddings on the WordNet Nouns Hypernymy Graph). Deep hyperbolic graph networks have never had such an audit. They are still compared on fixed splits, with hand-set baselines, and with geometry, optimizer, readout and auxiliary losses all changing at once.

We propose to measure what curvature alone contributes. The method uses a single implementation in which Euclidean space is just the curvature-zero setting of the same model. The experiments vary curvature while holding the code, the data splits and the tuning budget fixed. In two sentences: write one graph network whose only geometric knob is a curvature parameter, tune every setting with the same budget over many random splits, and see which reported hyperbolic gains survive. Then add back the other ingredients that hyperbolic papers bundle in, one at a time, to learn which of them actually carries the gain.

Core Claim

Our hypothesis is that much of the reported advantage of hyperbolic graph networks on citation graphs comes from training and evaluation choices rather than from geometry. A real, measurable advantage should remain only where theory predicts one: tree-like graphs at low embedding dimension. When Curvature Beats Dimension proves the relevant bound. A balanced $b$-ary tree of height $h$ embedded in $d$ Euclidean dimensions has distortion that grows like $(b^{\lfloor h/2\rfloor})^{1/d}$, while the hyperbolic plane achieves constant distortion. The proposal therefore makes a falsifiable prediction. Under a fair protocol, the curvature effect should shrink as $d$ grows and as the graph becomes less tree-like. Tree-likeness is measured by Gromov $\delta$-hyperbolicity, a number that is small when every triangle in the graph is thin, as in a tree.

What is new is the instrument, not one more benchmark run. Are Hyperbolic Representations in Graphs Created Equal? supplies the method. Its κ-stereographic model, with Taylor expansions near zero curvature, lets curvature $\kappa$ pass smoothly through $0$, so one codebase covers spherical, flat and hyperbolic geometry. At $\kappa=0$ the Möbius operations reduce to ordinary vector operations, and Riemannian Adam reduces to Adam. A comparison of $\kappa<0$ against $\kappa=0$ therefore removes implementation differences of the kind that produced the WordNet artifact. That paper itself used fixed splits and did not ask whether its main finding was a tuning artifact. Its finding was that Euclidean models win node classification, reaching about 0.809 on Cora, while hyperbolic models win link prediction. We test exactly that question.

Research Program

The first component is a geometry-fair tuning protocol, which addresses the open problem that search spaces cannot be compared across geometries. On the State of the Art of Evaluation in Neural Language Models and Fantastic Pretraining Optimizers and Where to Find Them motivate the general rule of giving each method an equal tuning budget. They did not face the further problem that learning-rate sensitivity differs sharply between flat and curved models, which Bansal and Benton show directly. We handle this in two steps. Each configuration first gets a short learning-rate range test that sets the center of its search interval, so no geometry inherits another's range. Each configuration then receives the same number of trials and the same GPU-hours. When curvature is a searched parameter, it uses up trials that the Euclidean model spends on other settings. We report full curves of best validation score against trials spent, not a single tuned number, so readers can see whether a conclusion depends on budget.

The second component is the evaluation protocol of Pitfalls of Graph Neural Network Evaluation, applied to hyperbolic models for the first time. It uses many random splits with several seeds each and identical early stopping. Shchur et al. used 100 splits with 20 seeds. We plan 50 splits with 5 seeds for the full factorial and 100 splits for the headline comparisons. We also add the feature-only multilayer perceptron baseline of A Fair Comparison of Graph Neural Networks for Graph Classification, adapted to node tasks with degree features.

The third component is a factorial decomposition of what hyperbolic papers change together. The factors are the following:

These factors map onto prior papers that each changed several of them at once, which is why their gains cannot be attributed. We fit an analysis of variance over the factors per task. The result is an attribution of the score gap to geometry, readout, losses and optimizer, together with their interactions.

A small shallow-embedding track on WordNet nouns and the YEAST network uses the same protocol. Its purpose is to settle the conflict between Bansal and Benton and Bridging ML and algorithms: comparison of hyperbolic embeddings. The latter reports high-dimensional Euclidean baselines underperforming low-dimensional hyperbolic ones and attributes this to optimization difficulties. The track also calibrates the protocol: it must reproduce Bansal and Benton's corrected numbers before we trust it on deep models.

Baselines

Naive reference points. The do-nothing reference is a feature-only multilayer perceptron with degree features, tuned under the identical budget. Errica et al. found that this kind of model beat every graph network they tested on D&D (78.4%) and ENZYMES (65.2%). Any geometric gain must first clear it. The second reference is a standard Euclidean GCN and GAT from widely used library code, tuned under the same budget. This is the strongest flat competitor, because Shchur et al. found that a well-tuned GCN reached 99.4% relative accuracy across datasets. Comparing it with our own $\kappa=0$ model measures the implementation tax: any accuracy lost purely because a model is written in hyperbolic-style code.

Strongest hyperbolic methods. The strongest existing hyperbolic methods are HGCN, run from its published design with learned per-layer curvature, and NGAT/NGCN from Khatir et al., with HIE added on top of HGCN. The reported figures to beat or reproduce are the following:

Method Dataset and task Reported result
HGCN DISEASE, link prediction 63.1% error reduction in ROC AUC over the best Euclidean GCN at 16 dimensions
HGCN AIRPORT, F1 Up to 47.5% error reduction
NGAT CiteSeer node classification 73.8%, against 71.6% for GAT and 68.1% for HGAT
HGCN + HIE CiteSeer node classification 74.1%, up from 67.6%

No reported result was found on Lacuna for HGCN's absolute PubMed number.

All methods share data splits, input features, early stopping, budget in trials and GPU-hours, and the range-test-centered search spaces. A hyperbolic method "beats" the tuned Euclidean model only under three conditions. Its mean over splits must exceed the Euclidean mean by more than the 95% interval of the paired split-wise difference. That margin must hold at every budget past the point where both curves flatten. And it must also hold against the same-code $\kappa=0$ variant.

Key ablation. The ablation that isolates our central ingredient is $\kappa<0$ against $\kappa=0$ within the same code, with every other factor fixed at its jointly tuned value.

Why not just…?

Why not just rerun HGCN against GCN over 100 random splits with an equal grid? That is our headline experiment, and it may by itself show the reported gaps shrinking. It cannot say why. HGCN and GCN differ in their projections, optimizer, curvature learning, attention aggregation and default hyperparameters. A surviving gap could come from any of these, and so could a vanishing one, through an HGCN-specific tuning failure. The same-code curvature sweep is what turns "the gap changed" into "curvature caused X of it."

Why not just use Euclidean layers with a hyperbolic readout, since Pseudo-Poincaré is faster and scored better? This is a real threat to the claim that hyperbolic message passing is needed. If the terminal-tanh variant matches full hyperbolic layers under fair tuning, that is a finding, and the factorial is designed to detect it. Khatir et al. themselves evaluated fixed configurations on standard splits without equal tuning. They also report vanishing gradients when the projection is stacked in deep models. Whether their 73.8% CiteSeer advantage holds up is therefore open, and the readout factor answers it directly.

Why not just raise the Euclidean dimension? Bansal and Benton show this closes the gap on WordNet by $d\ge 50$. The tree distortion bound says it cannot close it at fixed low $d$ on deep trees. The dimension sweep, run on synthetic trees of controlled depth and branching factor, tests both statements. If high-dimensional Euclidean models match hyperbolic ones on DISEASE and AIRPORT as well, the practical case for hyperbolic graph networks reduces to memory-constrained settings, and we will say so.

Why not just use fast combinatorial embedders such as BFKL? For shallow reconstruction they are a strong competitor: 0.756 MAP on YEAST against 0.542 for Lorentz embeddings in Celińska-Kopczyńska and Kopczyński. We include BFKL in the shallow track. They do not use node features, so they cannot address semi-supervised node classification, which is the main setting here.

Proposed Experiments and Decisive Checks

Benchmarks and metrics. We use DISEASE, AIRPORT, Cora, CiteSeer and PubMed for link prediction (ROC AUC and average precision) and node classification (accuracy and F1). We add Barabási-Albert graphs and the TREE-L/TREE-H synthetic trees, where depth, branching factor and homophily are controlled. The shallow track uses WordNet nouns and YEAST, scored by MAP and mean rank. Each graph's δ-hyperbolicity is recorded as a covariate.

Check 1: calibration. The protocol must reproduce Euclidean ≈0.92 MAP at 200 dimensions on WordNet and a hyperbolic advantage below 20 dimensions. Otherwise the deep-model results are not interpreted.

Check 2: the curvature effect. For each dataset and dimension, we compare $\kappa<0$ against $\kappa=0$ in the same code under equal budget. The prediction is a significant curvature effect on DISEASE and the synthetic trees at $d\le 16$, decaying with $d$, and an effect statistically indistinguishable from zero on Cora, CiteSeer and PubMed node classification. If curvature still loses on Cora node classification after Riemannian tuning, Kochurov et al.'s task split reflects geometry. If the loss disappears, it was a tuning artifact.

Check 3: predicting when curvature helps. We regress the size of the curvature effect across all graphs on δ-hyperbolicity and dimension. We then test whether the closed-form decision rule of When Curvature Beats Dimension, below 7% misclassification on its synthetic trees, predicts which real graphs benefit.

Check 4: attribution. The analysis of variance gives the share of the published hyperbolic-versus-Euclidean gap explained by each factor: geometry, readout, auxiliary losses, optimizer and tuning budget.

Releases and reading the results. We will release the tuning-curve data and the single codebase, so later hyperbolic papers can report their numbers on the same footing. A clean result in either direction is useful. Geometry may survive only where theory predicts it, or it may not survive at all.