Fair representation learning compresses data into an internal code that predicts a task well while revealing as little as possible about a protected attribute such as gender, age or region. The strongest version of this idea measures that leakage as mutual information, in bits, between the code and the attribute. A low enough value limits how unfair any downstream classifier can be. Separately, hyperbolic embeddings place data in a curved space where room grows exponentially with distance from the center. That makes them a natural fit for trees and social networks: general or central items sit near the origin and specific or peripheral items sit near the edge. The two ideas have never been combined.
Our central observation is that the combination is useful for one specific reason. A hyperbolic code has a built-in, meaningful split into depth (distance from the origin, which encodes centrality or hierarchy level) and direction (which branch or community a point belongs to). Total leakage then divides exactly into a depth part and a direction part. We propose to measure and budget the two parts separately. On social graphs, demographic groups often sit at different structural positions. If much of the protected information flows through depth, which is also task-relevant, then a single scrubbing budget destroys hierarchy that a practitioner might choose to keep, audit, or condition on. A split budget makes that trade-off explicit and keeps an overall guarantee on total leakage.
Core Claim
Let $z$ be a node's code on the Lorentz hyperboloid. Let $R = d(0, z)$ be its hyperbolic distance from the origin, and let $\Theta$ be the unit direction of its image in the flat tangent space at the origin. Because the map $z \mapsto (R, \Theta)$ is invertible away from the origin, the chain rule of mutual information gives an exact identity:
$$I(Z;S) = I(R;S) + I(\Theta;S \mid R),$$
where $S$ is the protected attribute. We train a hyperbolic graph encoder to maximize task utility subject to two separate budgets, $I(R;S) \le \epsilon_R$ and $I(\Theta;S \mid R) \le \epsilon_A$. We enforce them with learned Lagrange multipliers. The total leakage stays bounded by $\epsilon_R + \epsilon_A$, so the guarantee against any downstream classifier from Controllable Guarantees for Fair Outcomes via Contrastive Information Estimation is preserved. Within that total, practitioners can see and choose how much leakage is carried by structural position and how much by community membership.
The claim is falsifiable in three ways, and all three must hold:
- In unconstrained hyperbolic embeddings of real social graphs, a substantial share of $I(Z;S)$ concentrates in $R$.
- Allocating the budget between depth and direction gives a better fairness–utility frontier than a single total budget at equal total leakage.
- Both effects are weaker when the same polar split is applied to a Euclidean code, where the norm carries no hierarchical meaning.
What Is Actually New, and a Correction to the Obvious Idea
The obvious project at this intersection is "derive a hyperbolic version of the contrastive log-ratio upper bound." We think that gap is largely illusory. Mutual information does not change under invertible reparameterization, so leakage from a hyperbolic code equals leakage from its image under the logarithm map at the origin, which is an ordinary vector. The Euclidean estimator from CLUB: A Contrastive Log-ratio Upper Bound of Mutual Information applies there unchanged, up to the usual caveat that it is a true upper bound only when its learned conditional model fits well. A "hyperbolic encoder plus CLUB on tangent coordinates" is therefore our baseline, not our contribution. Geometry cannot change how much information the code holds. It can change how that information is organized, and that is what we exploit.
The prior work plays three roles here.
Work that motivates the proposal. Rethinking Fair Representation Learning for Performance-Sensitive Tasks shows that when the protected attribute and the label are genuinely correlated, full scrubbing unavoidably costs utility. FD-VAE (Learning Disentangled Representation for Fair Facial Attribute Classification via Fairness-aware Information Alignment) shows that splitting the code into target, protected and shared subspaces can recover some of that utility. Its subspaces are learned coordinate slices in a flat space. Ours is a geometric split that hyperbolic training produces without supervision, as the depth-as-generality behavior in Poincaré Embeddings for Learning Hierarchical Representations and Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry demonstrates. The Enron result in the latter, where executives cluster at the origin purely from communication patterns, is exactly the situation in which a protected attribute correlated with seniority would leak through depth. Learning Fair Graph Representations with Multi-view Information Bottleneck (FairMIB) shows that topology alone, with node features neutralized, regenerates sensitive information. That is direct evidence that structural position is a major leakage channel.
Work that supplies the method. We reuse three pieces:
- the Lagrangian budget mechanism of L-MIFR (Learning Controllable Fair Representations), which adjusts penalty weights automatically until a user's bit budget is met;
- the decoder-free contrastive leakage bound and utility term of FCRL;
- hyperbolic message-passing encoders from A Hyperbolic-to-Hyperbolic Graph Convolutional Network and Hyperbolic Graph Neural Networks, run in the Lorentz model for numerical stability near the boundary.
Work that constrains the design. FCRL's argument that adversarial training only lowers a lower bound on leakage rules out a gradient-reversal adversary as the fairness mechanism. FairMIB itself, the strongest graph fairness method we know of, works entirely in Euclidean space and has no notion of depth versus direction.
Research Program
The encoder is a Lorentz-model hyperbolic graph network that produces $z$ for each node. Training is Riemannian stochastic gradient descent for manifold parameters and ordinary optimizers for the rest.
Estimating the depth term. $I(R;S)$ involves a one-dimensional variable and a discrete attribute. It can therefore be estimated with low-variance plug-in methods (binned or kernel density estimates of $p(R \mid S)$). This is a much easier problem than estimating high-dimensional leakage, and it gives a check on the variational bounds.
Estimating the direction term. $I(\Theta;S \mid R)$ uses a CLUB-style bound with a conditional model of $S$ given $\Theta$ and $R$.
Diagnostic stage. We train unconstrained hyperbolic and Euclidean encoders, then report what fraction of total leakage sits in depth versus direction. Total leakage is cross-checked by the strongest post-hoc probe we can train, for example a wide multilayer perceptron predicting $S$ from $z$. This stage answers the question of where bias lives, and it is publishable on its own if the answer is surprising in either direction.
Constrained stage. We train under split budgets across a grid of $(\epsilon_R, \epsilon_A)$ pairs and trace the fairness–utility frontier at fixed total budget. Reporting is staged:
- conditional demographic parity within depth bins, which is what a direction-only constraint controls;
- ordinary demographic parity and equalized odds, which the total budget controls;
- an honest note that allocating budget to depth is a normative choice. It is appropriate when centrality is a legitimate factor and inappropriate when centrality is itself a product of discrimination.
Baselines
The naive reference points are unconstrained encoders: a standard Euclidean graph convolutional network and an unconstrained hyperbolic graph network. Both are trained only for the task on the same splits. Next comes fairness through unawareness, in which the protected attribute is removed from node features before hyperbolic training. FairMIB shows topology alone still leaks, so this should fail, but it sets the floor.
FairMIB. This is the strongest existing competitor on graphs because it already targets multiple leakage channels, including structure. The file reports that it "reduced fairness violations (DP and EO) by over 98% compared to a standard GCN" on German. It also "improved the F1-score by 3.2% while simultaneously reducing bias" on Bail. For Pokec we found no reported numerical result on Lacuna. We will rerun FairMIB from its released configuration on Bail, Credit, German, Pokec-z and Pokec-n with our splits. It gets the same hyperparameter search budget, measured in trials, as our method and an equal embedding dimension sweep.
Euclidean contrastive bound. The strongest information-theoretic competitor is FCRL-style training with a single total leakage budget on a Euclidean graph encoder. On Adult, Controllable Guarantees for Fair Outcomes via Contrastive Information Estimation reports an area score of 0.307 for FCRL versus 0.251 for MIFR. We use FCRL's frontier-area metric on all datasets so the comparison is in its own terms.
Ablations. Two ablations isolate our ingredients:
- Hyperbolic, total budget only: the "hyperbolic encoder plus CLUB on tangent coordinates" model with one budget on $I(Z;S)$. This isolates the value of splitting.
- Euclidean, polar split: the identical split budget applied to norm and direction of a Euclidean code. This isolates the value of curvature.
What counts as winning. Beating the baselines means two things. First, at matched measured total leakage, confirmed by post-hoc probes and not only by our own bound, our frontier area exceeds both the hyperbolic total-budget ablation and the Euclidean FCRL baseline on at least the Pokec graphs. Second, the gain over the Euclidean polar ablation is clearly larger than seed variance. Matching FairMIB's demographic parity and equalized odds reductions on German while keeping comparable utility at lower dimension is the minimum bar against the strongest system.
Why not just…?
Why not just use a Euclidean fair encoder with more dimensions? On hierarchies, Poincaré Embeddings for Learning Hierarchical Representations reports a mean average precision of 0.823 for a 5-dimensional hyperbolic embedding versus 0.168 for a 200-dimensional Euclidean one, so raw capacity is not a substitute. That evidence concerns reconstruction, not fairness, so it does not settle our question. The dimension sweep in our Euclidean baselines tests it directly. If a 256-dimensional Euclidean FCRL matches our frontier, the geometric argument fails on these graphs, and we will report that.
Why not just condition on node degree in a Euclidean model? This is the most serious threat. A Euclidean encoder constrained on $I(Z;S \mid \text{degree})$ captures part of the depth-versus-direction idea with no curved geometry at all. Our answer is that hyperbolic depth is a learned, multi-hop notion of hierarchical position rather than a raw count, but that is a hypothesis, not a fact. We add this degree-conditioned Euclidean model, and a PageRank-conditioned variant, as baselines. If they match the hyperbolic split, the contribution reduces to the diagnostic finding and the budget-splitting procedure, and the geometry is unnecessary.
Why not just train a hyperbolic embedding and add an adversary or post-hoc threshold? FCRL argues that adversaries minimize a lower bound on leakage, so a weak adversary can hide leakage that a stronger downstream model recovers. In FD-VAE's CelebA experiments, adversarial training left an equal-opportunity violation of 7.6 against 1.3 for information alignment. Post-hoc thresholds also do nothing for downstream users who train new classifiers on the code. We run the adversarial hyperbolic variant and evaluate it with probes stronger than its training adversary to show the gap.
Why not just remove the protected attribute before training? FairMIB's structural-view result shows that topology alone reconstructs sensitive information, and our unawareness baseline will quantify this for hyperbolic encoders specifically. We expect this leakage to sit largely in depth, which would itself support the core claim.
Proposed Experiments
Datasets. The primary testbed is the graph fairness benchmarks used by FairMIB: Bail, Credit, German, Pokec-z and Pokec-n. Pokec is the most plausible place for hierarchical, power-law structure to matter. Adult and Heritage Health serve as negative controls, because tabular data without hierarchy should show no benefit from curvature. A hyperbolic advantage there would signal an artifact rather than a geometric effect.
Metrics.
- Task AUC and F1.
- Demographic parity and equalized-odds gaps.
- Conditional parity within depth bins.
- Estimated $I(R;S)$, $I(\Theta;S \mid R)$ and their sum, cross-checked by probe-based leakage.
- Frontier area in FCRL's sense.
- Whether measured leakage tracks the requested budgets, following L-MIFR's controllability test.
Decisive checks. There are three, in order:
- Localization. On Pokec, does the depth term carry a large, task-correlated share of leakage in hyperbolic codes but not in Euclidean norms? If not, the project stops at a clean negative result.
- Allocation. At equal total budget, does shifting budget toward depth recover utility relative to the total-budget hyperbolic ablation, while probe-measured total leakage stays within $\epsilon_R + \epsilon_A$?
- Curvature and centrality. Does the hyperbolic split beat both the Euclidean polar split and the degree-conditioned Euclidean model? Passing the first two checks but failing this one would mean that splitting the budget is valuable but hyperbolic geometry is not. We would present the method that way.
References
- Controllable Guarantees for Fair Outcomes via Contrastive Information Estimation
- CLUB: A Contrastive Log-ratio Upper Bound of Mutual Information
- Rethinking Fair Representation Learning for Performance-Sensitive Tasks
- Learning Disentangled Representation for Fair Facial Attribute Classification via Fairness-aware Information Alignment
- Poincaré Embeddings for Learning Hierarchical Representations
- Learning Continuous Hierarchies in the Lorentz Model of Hyperbolic Geometry
- Learning Fair Graph Representations with Multi-view Information Bottleneck
- Learning Controllable Fair Representations
- A Hyperbolic-to-Hyperbolic Graph Convolutional Network
- Hyperbolic Graph Neural Networks.