Fairness interventions often reach group parity the cheap way: they make the best-served group worse rather than making the worst-served group better. This is the "leveling-down" effect. We propose to diagnose and fix it with a geometric signal that only exists in hyperbolic space. Hyperbolic space is a curved space whose available room grows exponentially with distance from a center point. When a classifier embeds its features there, an example's distance from the center naturally acts as a readout of how specific and well-resolved the model's representation of it is. General, poorly resolved content sits near the center; confident, specific content sits near the edge. Our hypothesis is that under-represented groups are pushed toward the center because they are underfit. Standard fairness fixes then close the gap mostly by pulling the majority inward too. The remedy we propose is one-sided. We push minority and intersectional subgroups outward to match the majority's depth, and we freeze the majority's radius. The exponential growth of room near the edge means the added minority mass should not crowd out majority structure.
The two-sentence version is this. Put subgroups (class × protected attribute, and their intersections) as children of their class in a low-dimensional hyperbolic classifier, and measure each group's average distance from the center. Then add a fairness penalty that only moves lagging groups outward. The test is whether this raises worst-group accuracy without the best-group drop that current methods still pay. A second test is whether the advantage over a flat-space version grows with the number of intersectional subgroups, as the geometry predicts.
Core Claim
Leveling-down has a measurable geometric signature, and hyperbolic geometry lets us repair it without trading one group against another. Concretely, we make three claims. First, in a hyperbolic classifier trained with plain average-loss minimization, worst-group examples have systematically smaller distance from the center than best-group examples. Second, existing reweighting methods close the accuracy gap partly by shrinking best-group radii. Third, a radial penalty that lifts only lagging groups outward, with best-group radii held fixed, recovers worst-group accuracy while keeping best-group accuracy at its unconstrained level.
The geometry-specific part of the claim is falsifiable. Take a fixed embedding dimension and grow the number of subgroups by crossing more attributes. The gap between the hyperbolic and flat-space versions of the same one-sided penalty should widen. The gap should shrink as the flat-space dimension grows. If neither happens, the benefit comes from the one-sided penalty rather than from curvature, and we will report it that way.
Positioning Against Prior Work
The motivating evidence comes from the fairness-accuracy line. BM-CL recasts leveling-down as catastrophic forgetting of the advantaged group during fairness fine-tuning, and anchors the network with Learning without Forgetting or Elastic Weight Consolidation (penalties that keep a network close to its earlier behavior). It shrinks, but does not remove, the best-group drop on CelebA. We share its diagnosis that the majority is being damaged. We differ in the fix: BM-CL anchors parameters, which live in shared linear subspaces, while we anchor a representation-level quantity, the majority's radius, and give the minority new room rather than competing room. Feature and Label Embedding Spaces Matter shows that bias can be isolated as a direction in a cosine-similarity feature space and projected out. That paper supplies the idea that bias has geometric form. It assumes one binary attribute and a linear bias direction; our hierarchy treats intersections as deeper levels instead. Bias-tolerant Fair Classification and FAE support the view that bias is often underfitting of minority groups rather than an intrinsic trade-off. That view is exactly what a radius-as-resolution readout would expose.
The method comes from the hyperbolic line. Poincaré Embeddings establish that hierarchies fit in two or three hyperbolic dimensions with little distortion. APo-VAE finds general content near the center and specific content at the edge, which is the basis of our radius readout. Hyperbolic Space with Hierarchical Margin shows that inward-curving shortest paths keep fine-grained classes from crowding under coarse labels. We reuse its parent-child prototype margin, with subgroups as the children. Clipped Hyperbolic Classifiers constrains the norms of hyperbolic features for stable training. That constraint matters here because our penalty pushes norms outward, so clipping sets the ceiling we lift toward.
The closest hyperbolic-and-imbalance work is ASIL. It shows hyperbolic graph clustering keeps minority clusters from being absorbed by majority ones. ASIL is unsupervised and has no protected attributes or group-fairness metrics. Addressing Fairness in Classification with a Model-Agnostic Multi-Objective Algorithm uses "hyperbolic" only in the sense of a tanh relaxation in flat space, so it is not a geometric predecessor.
Research Program
The model is a standard image backbone followed by an exponential map into a Poincaré ball: the open unit ball with a metric that stretches distances as points approach the boundary. The ball has low dimension (8 to 32), with feature clipping as in Clipped Hyperbolic Classifiers. Classification uses hyperbolic distance to learned prototypes. Each class prototype has child prototypes for each subgroup, tied by the hierarchical margin of Hyperbolic Space with Hierarchical Margin. For example, on Waterbirds these are waterbird-on-land and waterbird-on-water. Intersectional subgroups sit one level deeper. The distance of feature $z$ from the center is
$$r(z) = 2\,\mathrm{artanh}(\|z\|),$$
where $\|z\|$ is its Euclidean norm inside the unit ball.
The first phase is diagnostic and needs no new method. We train plain ERM and GroupDRO hyperbolic models and plot per-group distributions of $r$ against per-group accuracy. For GroupDRO we also track how the best group's mean radius changes relative to ERM. The hypothesis fails if minority radii are not smaller under ERM, or if the best group's radius does not fall under GroupDRO. In that case the rest of the program loses its rationale, and we would stop.
The second phase adds the one-sided radial penalty to the classification loss $\mathcal{L}_{\text{cls}}$:
$$\mathcal{L} = \mathcal{L}_{\text{cls}} + \lambda \sum_{g} \max\big(0,\ \mathrm{sg}[\bar r_{\text{ref}}] - \bar r_g\big) + \mu\,\big(\bar r_{\text{best}} - \mathrm{sg}[\bar r^{\,0}_{\text{best}}]\big)^2 .$$
The terms are defined as follows. $\bar r_g$ is the mean radius of correctly labeled training examples in group $g$ within a batch. $\bar r_{\text{ref}}$ is the mean radius of the best-served group. $\mathrm{sg}$ stops gradients, so lagging groups move and the reference does not. $\bar r^{\,0}_{\text{best}}$ is the best group's radius from the ERM checkpoint. $\lambda$ and $\mu$ weight the two penalties. The last term replaces BM-CL's parameter anchor with a radius anchor.
When group labels are missing for intersections, we group by predicted subgroup prototype. This is the route we would use to reach settings like Safe Fairness Guarantees Without Demographics in later work.
Baselines
The naive reference point is plain ERM with the same backbone and a Euclidean linear head. It sets the unconstrained best-group accuracy that any method should not fall below. Beating it means higher worst-group accuracy with a best-group change within seed noise.
The strongest existing method on this exact question is BM-CL, in its ReSample-LwF variant. It reports Waterbirds advantaged-group accuracy held at 99.3% and worst-group accuracy lifted from 72.8% to nearly 80%. On CelebA it reports a best-group accuracy drop of 2.8% versus 3.5% for GroupDRO, and on CheXpert global and balanced accuracy of around 77%. We will rerun it ourselves on identical splits, backbone, epochs, and a matched hyperparameter search of the same number of trials, and quote its paper numbers only as a sanity check. Beating it means worst-group accuracy at least as high with a CelebA best-group drop clearly below 2.8%. Ideally that drop is statistically indistinguishable from zero over five seeds.
GroupDRO, the reweighting method that minimizes the worst group's loss, is run in both Euclidean and hyperbolic heads to show where leveling-down originates. Feature and Label Embedding Spaces Matter is included on CelebA as the geometric-debiasing competitor. It reports CelebA fairness metrics (Bias −0.041, Parity 6.71, Opp 6.73, Odds 4.10) but no absolute mAP; we will report its accuracy from our rerun.
The key ablations isolate each ingredient. A Euclidean cosine head with the identical one-sided penalty on feature norm isolates curvature. A hyperbolic head with no radial penalty isolates the loss. A symmetric radial penalty that equalizes all radii isolates one-sidedness, since it should reproduce leveling-down. A hyperbolic head without the subgroup hierarchy isolates the prototype tree.
Why not just…?
Why not just make the Euclidean embedding bigger? Flat-space volume grows only polynomially with radius, so resolving many subgroups needs many dimensions. Poincaré Embeddings fit taxonomies in two or three hyperbolic dimensions. There is no reported result on Lacuna for dimension sweeps in group-fairness classification, so this is untested here. Our dimension sweep (8 to 1024 for the flat version) settles it. If flat space at moderate dimension matches us at every subgroup count, the curvature claim is wrong.
Why not just anchor the network with LwF or EWC, as BM-CL does? It is the strongest competitor and already nearly preserves Waterbirds best-group accuracy. It still pays 2.8% on CelebA, and BM-CL reports sensitivity to its regularization weights. We think that residual cost comes from anchoring parameters that both groups share. Our radius anchor constrains one number per group instead. Head-to-head CelebA and CheXpert runs with matched tuning budgets decide this. We also combine LwF with our loss to check whether the two effects add.
Why not just apply the same one-sided norm penalty in Euclidean space? This is the most serious threat, because the one-sidedness might do all the work. The Euclidean-norm ablation is the direct test. The geometry predicts a specific pattern, not just a win: at fixed dimension, the hyperbolic advantage should grow as we cross more attributes on CelebA (sex, then sex × age, then sex × age × hair color). Flat space runs out of room faster there. A constant or absent gap would mean the contribution is a Euclidean norm-fairness loss, and we would present it as such.
Proposed Experiments
We run Waterbirds, CelebA, and CheXpert with the group definitions used by BM-CL. We add an intersectional CelebA stress test with 2, 4, and 8 subgroups per class, and Adult with sex × race as a tabular check where no absolute prior numbers are reported on Lacuna. Metrics are worst-group accuracy, best-group accuracy change relative to ERM (our direct measure of leveling-down), average and balanced accuracy, and equalized-odds gap. Each is reported over five seeds.
Four checks are decisive. First, the diagnostic radius gap must exist under ERM and must shrink under GroupDRO partly through a falling best-group radius. Second, the one-sided hyperbolic loss must beat rerun BM-CL on CelebA best-group change at equal or better worst-group accuracy. Third, the symmetric radial variant must reproduce leveling-down, which confirms the mechanism. Fourth, the hyperbolic-minus-Euclidean gap must widen with subgroup count at fixed dimension and shrink as flat-space dimension grows. If the first check fails we stop. If the second and third pass but the fourth fails, the finding is a useful flat-space fairness loss rather than a geometric one.
References
- BM-CL: Bias Mitigation through the lens of Continual Learning
- Feature and Label Embedding Spaces Matter in Addressing Image Classifier Bias
- Bias-tolerant Fair Classification
- FAE: A Fairness-Aware Ensemble Framework
- Poincaré Embeddings for Learning Hierarchical Representations
- APo-VAE: Text Generation in Hyperbolic Space
- Hyperbolic Space with Hierarchical Margin Boosts Fine-Grained Learning from Coarse Labels
- Clipped Hyperbolic Classifiers Are Super-Hyperbolic Classifiers
- ASIL : Augmented Structural Information Learning for Deep Graph Clustering in Hyperbolic Space
- Addressing Fairness in Classification with a Model-Agnostic Multi-Objective Algorithm
- Safe Fairness Guarantees Without Demographics in Classification: Spectral Uncertainty Set Perspective