Item Radius as an Exposure Dial in Hyperbolic Recommenders

Recommender systems that embed users and items in hyperbolic space place popular, general items near the center and niche items near the edge. Hyperbolic space is a curved geometry whose volume grows exponentially with distance from the center, much like the number of nodes in a tree grows with depth. In a recommender that scores items by distance to the user, an item near the center is moderately close to almost everyone, so it lands in many users' top slots. That makes an item's distance from the center, its radius, a de facto control over how much user attention it receives. The radius is learned mostly from interaction counts, and those counts are themselves produced by past rankings. So the geometry quietly encodes the rich-get-richer loop that the fair-ranking literature worries about: items shown often get clicked, get pulled inward, and get shown more.

We propose to make that control explicit and use it. Each item keeps its direction, which says which users it suits. Top-$K$ exposure fairness, meaning equitable attention across item groups in the slots users actually look at, is enforced by adjusting only item radii, which say how broadly the item is shown. The bet is that separating "who sees it" from "how many see it" lets a fairness objective cut group exposure disparity at lower accuracy cost than applying the same objective to all embedding coordinates. It should also cost less than the simplest Euclidean counterpart, a per-item score offset. The comparison with that offset decides whether the curvature matters or only the decomposition does.

Core Claim

In a distance-based hyperbolic recommender, write user $u$ and item $i$ by their radii $r_u, r_i$ and the angle $\theta$ between their directions. For points well away from the center and not nearly aligned, the hyperbolic distance is approximately

$$ d(u,i) \approx r_u + r_i + 2\log\sin(\theta/2), $$

which follows from the hyperbolic law of cosines when $\cosh$ is replaced by its exponential leading term. For nearly aligned points (same branch of the implicit tree), the distance instead behaves like $|r_u - r_i|$. Moving an item radially therefore does two different things. For users on other branches it acts like a uniform score offset. For users on the item's own branch, the specialists who sit deeper than it, moving it inward pushes it away. An inward move trades specialist relevance for broad reach, and an outward move does the reverse.

This matters for fairness because much of a head item's exposure comes from cross-branch users who see it only because it is central. Pushing majority-group head items outward removes exactly that popularity-driven exposure while keeping them close to their actual audience. An additive score offset cannot do this, since it demotes an item for fans and bystanders alike. Our claim is that constraining fairness pressure to the radial coordinate, inside a top-$K$ exposure objective, yields a better accuracy–disparity trade-off than the obvious alternatives on sparse, tree-like catalogs. We also predict the advantage vanishes on dense catalogs, where hyperbolic geometry itself stops helping.

Position in the Literature

Two literatures meet here without touching. The fair-ranking side supplies the objective. Fairness of Exposure in Rankings defines exposure as rank-discounted attention and optimizes doubly stochastic ranking matrices per query. It also shows the motivating winner-take-all effect: a 0.03 difference in average relevance produced a 0.32 difference in average exposure. Policy Learning for Fairness in Ranking learns a stochastic Plackett–Luce ranking policy penalized by exposure disparity. Learning to Rank with Top-$K$ Fairness (KSO-RED) shows that enforcing exposure balance over a whole long list wastes the fairness budget on positions nobody inspects. It restricts disparity to the top $K$ through a bilevel formulation with a threshold on the $(K{+}1)$-th score and moving-average gradient estimators. We reuse KSO-RED's top-$K$ exposure estimator as our fairness term. The difference is where its gradient is allowed to go.

All three operate on generic Euclidean scores with no notion of item geometry. We keep their exposure definitions and change only the parameter that absorbs fairness pressure. The per-query matrix methods also scale poorly to large catalogs. A per-item radius is one scalar per item, shared across all users, so it can also be re-solved after training without per-query matching.

The hyperbolic side supplies the model and a warning. Poincaré Embeddings for Learning Hierarchical Representations established that radius encodes generality and that trees embed with low distortion in few dimensions. Where are we in embedding spaces? shows that hyperbolic geometry helps distance-based recommenders, not inner-product ones, and mainly on sparse data at low dimension. On dense MovieLens, Euclidean and hyperbolic performance were comparable. That result constrains us: our mechanism only exists for distance scoring, and we should expect little on dense data.

HRCF already manipulates radii. It centers embeddings and adds a penalty pushing items away from the origin, reporting an 11.5% tail-item Recall@10 gain on Amazon-CD. Hgformer reports a 55.7% tail Recall@10 gain on Amazon-CD. Both treat tail performance as an accuracy problem and never measure how attention is distributed across ranked slots. HRCF's radial push is a uniform heuristic that is not tied to any exposure target. We turn radius into a variable optimized against a measured group exposure constraint.

One consequence we test directly: better tail recall does not imply fairer exposure. A model can retrieve tail items more accurately for the few users who want them while still concentrating top-slot attention on central items for everyone else.

Research Program

The backbone is an HRCF-style Lorentz-model collaborative filter with distance scoring. We use the Lorentz model, a numerically stable coordinate system for hyperbolic space. Each item is re-parameterized as a radius $r_i$ and a unit direction $\mu_i$, and the embedding is obtained by the exponential map from the origin along $r_i \mu_i$. Users are unchanged. This re-parameterization does not change what the model can express; it only exposes the radius as its own parameter.

The first step is an audit. On trained HGCF-, HRCF- and Hgformer-style models and on Euclidean LightGCN and CML, we measure each item's aggregate top-$K$ exposure. We use the logarithmic discount $1/\log_2(1+\text{rank})$ averaged over users, the discount that A Common Approach for Consumer and Provider Fairness in Recommendations argues is essential. We then ask three things. How much of the variance in exposure does radius explain, compared with interaction count? Is group exposure disparity larger or smaller in the hyperbolic models than in the Euclidean ones, even where tail recall is higher? What share of each item's exposure comes from same-branch versus cross-branch users? We define branches by angular clustering of user directions.

The second step is the method. The relevance loss (the backbone's usual margin ranking loss) updates both radius and direction. A KSO-RED top-$K$ exposure-disparity term, weighted by $\lambda$, updates radii only, with a stop-gradient on directions and user embeddings. We implement both demographic-parity and relevance-proportional (disparate-treatment) targets from Singh and Joachims.

The third step is a post-hoc variant. It freezes a trained model and solves for radii alone to meet a target disparity, one scalar per item. This tests whether exposure can be re-balanced at serving time without retraining.

Baselines

The naive reference is the unconstrained backbone, HRCF, alongside unconstrained Euclidean LightGCN and CML at matched dimension. These fix the zero-fairness end of every trade-off curve and establish the audit's claim about how exposure is distributed.

The strongest existing competitor is KSO-RED from Learning to Rank with Top-$K$ Fairness, because it targets exactly our fairness quantity: top-$K$ exposure disparity traded against NDCG. We run it two ways. First, as published on a Euclidean scorer. Second, and more stringently, with the identical top-$K$ disparity term on the same hyperbolic backbone but with gradients flowing to all coordinates. The second version is our key ablation: it isolates whether restricting fairness pressure to the radius matters. The file reports that KSO-RED dominates DELTR and a robust baseline on ML-20M and Netflix-20M Pareto curves. No reported result was found on Lacuna for KSO-RED's numeric NDCG or disparity values, so we reproduce it rather than quote it.

For scale reference on the stochastic-policy line, Fairness for Robust Learning to Rank reports NDCG 0.839 at fairness violation 0.031 on Adult, against 0.721 for Fair-PG-Rank at similar fairness. That is a small-candidate-set benchmark we do not reuse. We include Fair-PG-Rank with a distance-based Plackett–Luce policy on the hyperbolic backbone as a further in-processing competitor.

The decisive competitor for the geometry question is a Euclidean CML model with a learned per-item additive bias. The same top-$K$ fairness term updates only those biases. This is the "one scalar per item" control without curvature. We also include per-query fair re-ranking of the unconstrained HRCF output using the linear program of Fairness of Exposure in Rankings, on the catalog sizes where it is tractable.

Every method gets the same data splits and group definitions, the same embedding dimensions (10, 32 and 64), the same number of $\lambda$ values swept to trace a Pareto front, and the same hyperparameter-search budget in trials. Beating a baseline means a strictly higher NDCG@10 at matched top-$K$ disparity over the middle of the disparity range, summarized as the area under the Pareto front, with confidence intervals over five seeds.

Why not just…?

Why not just add the fairness loss to the whole hyperbolic model? This is the obvious combination and the most direct test of our idea, so it is our main ablation rather than something we argue against. If full-gradient KSO-RED on HRCF matches radius-only training at every disparity level, the decomposition buys nothing and we would report that. Our prediction that it will not rests on the audit. If disparity lives mostly in radius, full-gradient fairness pressure will also bend directions, scrambling taste structure and costing NDCG for specialist users. We measure this by splitting the NDCG loss into same-branch and cross-branch users.

Why not just use a per-item score offset in a Euclidean model? This is the real threat. By the distance approximation above, a radial move is an offset for cross-branch users, so any advantage must come from the second-order effect on same-branch users. If the Euclidean bias-only baseline ties radius-only hyperbolic training on the sparse Amazon datasets, the honest conclusion is that the gain comes from parameterizing a per-item exposure scalar, not from curvature. That would still be a useful, cheap recipe. We will know from the Pareto comparison and from whether the specialist-retention gap appears.

Why not just re-rank a Euclidean model's output under exposure constraints? Post-processing is exact for the constraint it enforces. However, it re-allocates exposure query by query from whatever scores the model produced. Policy Learning for Fairness in Ranking argues that in-processing reaches better trade-offs when the learned representation itself is biased. Per-query matrix solving also scales poorly. We include it where tractable. If it dominates, our method's case reduces to scalability, which the post-hoc radius variant addresses directly.

Why not just use HRCF's origin penalty or a larger Euclidean dimension? HRCF's penalty pushes all items outward regardless of group or exposure, so we expect it to move tail recall without closing group disparity. The audit and baselines check this directly. Larger Euclidean dimension closes the accuracy gap, as Where are we in embedding spaces? found between $d=10$ and $d=100$. It says nothing about exposure, so we run the Euclidean baselines at $d=64$ as well as low dimension.

Proposed Experiments

The fairness benchmarks follow KSO-RED. ML-20M-H groups horror versus non-horror films, ML-20M-D documentary versus non-documentary, and Netflix-20M pre-1990 versus later releases. To test the density prediction we add sparse catalogs from the hyperbolic literature, Amazon-CD and Amazon-Book, where we define groups two ways. One uses item category; the other uses Hgformer's popularity split, bottom 80% versus top 20%.

The metrics are NDCG@10 and Recall@10, top-$K$ exposure disparity under the log discount at $K\in\{10,20\}$, and exposure Gini across items. The mechanism-level measures are the correlation between radius and exposure, and the NDCG change split between same-branch and cross-branch users.

The experiments stand or fall on four checks. First, the audit must show that radius explains a large share of exposure in hyperbolic models. If it does not, the radius is not an exposure control and the project stops there. Second, radius-only training must beat full-gradient KSO-RED on the hyperbolic backbone in Pareto area on the sparse Amazon datasets. Third, it must beat the Euclidean bias-only baseline there, with the gain concentrated in retained NDCG for same-branch users. That is the signature of the non-uniform radial effect. Fourth, the advantage over Euclidean baselines should shrink on ML-20M, mirroring the density finding of Sun et al. If it does not shrink, the effect comes from something other than the tree-like geometry, and we would need a different explanation.

A final check tests whether post-hoc radius solving on a frozen model recovers most of the in-training Pareto front. That determines whether exposure fairness in hyperbolic recommenders can be adjusted at serving time at the cost of one scalar per item.