Recommender systems usually store each user as one fixed vector, even though tastes shift over weeks, sessions, and single evenings. A separate line of work shows that hierarchical catalogs are captured better in hyperbolic space than in flat space. Hyperbolic space is a curved geometry whose volume grows exponentially with radius, so a tree such as genre → artist → track fits into it without crowding. The two ideas rarely meet in streaming use because curved-space operations are slow. Every update of a hyperbolic state usually maps the point to a flat local approximation of the manifold and back again, which requires hyperbolic sines, cosines and square roots. GPUs execute these poorly. This proposal claims that the particular operation drift tracking needs, a recency-weighted average of what the user recently engaged with, can be computed on the hyperboloid (the Lorentz model of hyperbolic space) using only multiply-adds.
The mechanism fits in two sentences. Keep a running, time-decayed sum of the user's item embeddings in the ambient space of the hyperboloid. Because the closed-form hyperbolic centroid is just that sum rescaled, the result is a linear recurrence with no exponential maps, logarithmic maps, or parallel transport, and ranking reduces to one sign-flipped dot product. Two further properties come for free. The size of the sum measures how spread out recent behavior is, so exploration and drift can be detected with no extra model. Running a fast and a slow copy separates transient detours from changes in the user's core profile.
Core Claim
Let item embeddings $x_i$ lie on the hyperboloid $\{x : \langle x, x\rangle_L = -1/c\}$. Here $\langle a,b\rangle_L = -a_0b_0 + \sum_{k\ge1} a_k b_k$ is the Minkowski inner product and $c>0$ is the curvature magnitude. A user's state after the $t$-th interaction, which occurs at time $\tau_t$ on item $i_t$, is
$$ S_t = \gamma_t^{\,\tau_t - \tau_{t-1}}\, S_{t-1} + x_{i_t}, \qquad \mu_t = \frac{S_t}{\sqrt{c}\,\sqrt{-\langle S_t, S_t\rangle_L}} . $$
In this equation $\gamma_t \in (0,1]$ is a per-step retention factor. Raising it to the elapsed time makes decay respect irregular gaps. The point $\mu_t$ is the weighted Lorentzian centroid of the history, which is a valid point on the hyperboloid because nonnegative sums of points on the upper sheet stay inside the light cone. Hyperbolic distance from the state to a candidate item is $\tfrac{1}{\sqrt c}\operatorname{arcosh}(-c\langle \mu_t, x\rangle_L)$. That function is monotone in $-\langle S_t, x\rangle_L$, so top-$k$ retrieval is an ordinary maximum-inner-product search with the first coordinate negated. No transcendental function is evaluated per candidate.
The unnormalized sum also carries information. Write $W_t$ for the total decayed weight in $S_t$. The dispersion ratio
$$ r_t = \frac{\sqrt{c}\,\sqrt{-\langle S_t,S_t\rangle_L}}{W_t} \;\ge\; 1 $$
equals 1 only when all recent items coincide. The inequality follows from the reverse Cauchy–Schwarz inequality for timelike vectors. For two items at geodesic distance $d$, the ratio equals $\cosh(\sqrt{c}\,d/2)$. It therefore grows exponentially with how far apart recent interests are, which makes it a much more sensitive exploration signal than its Euclidean counterpart.
The model keeps a fast state and a slow state, each an instance of the recurrence above. A small learned gate reads $r_t$ and the agreement $-c\langle\mu^{\text{fast}}_t,\mu^{\text{slow}}_t\rangle_L$ between the two states, and it outputs both $\gamma_t$ and the mixing weight used for scoring. When the fast state disagrees with the slow one but is tightly concentrated, the gate should read the pattern as genuine drift and shorten the slow memory. When the fast state is diffuse, the gate should read it as a detour and protect the slow memory.
The novelty lies in the combination of three properties. The first is an update whose cost matches Euclidean hardware operations. The second is intrinsic hyperbolic geometry at both the state and the scoring layer. The third is a geometry-derived signal that sets the forgetting rate. None of the connecting works in the file has all three.
Relation to Prior Work
Several works motivate the proposal by showing the costs of curved-space layers. Hyperbolic Neural Networks builds network layers by mapping into flat tangent spaces and back, and Unsupervised Hyperbolic Representation Learning via Message Passing Auto-Encoders reports that this repeated mapping inflates wall-clock time and GPU demand. Riemannian Continuous Normalizing Flows notes that geodesic operations defeat the fused multiply-add pipelines of tensor cores. Our design avoids tangent maps altogether in the streaming path rather than making them cheaper.
The closest existing systems each give up one of the three properties above. HMamba runs a selective state-space recurrence in hyperbolic space. Its full variant pays extra overhead for hyperbolic operations, and its "half" variant regains speed by scoring in Euclidean space. Its state kernel also treats history as stationary.
SINCERE shows that the hyperbolicity of user interaction patterns drifts over time while item structure stays stable. This is the empirical motivation for letting user states move while item embeddings stay fixed. SINCERE, however, pays for tangent-space cross-geometry aggregation and for Ollivier–Ricci curvature estimation, which solves optimal-transport problems at each step.
Fast Mixture of Curvature-Aware Experts is the methodological precedent for removing exponential and logarithmic maps by algebraic reformulation. It rewrites Riemannian attention using polar decomposition and the law of cosines, reporting a 27.5% speedup and a 52.6% memory reduction. That work targets dynamic-graph link prediction and does not model user drift. We apply the same move to a different operation, the weighted mean, and get a linear recurrence instead of cheaper attention. H²SeqRec and THINK combine hyperbolic embeddings with time. The first bins time into months, quarters and years, and the second uses fixed-curvature convolutions. Neither addresses continuous drift or latency.
On the drift side, Right Now, Wrong Then (NS-DPO) supplies the exponential time-discounting template. It also constrains the design, because the authors acknowledge that the discount must be tuned against an unknown variation budget. Our dispersion-driven gate is a direct attempt to remove that tuning. Walk2Pers splits preference moves into magnitude and orientation and keeps two memory lanes, but it works in flat space with fixed decay kernels that the authors acknowledge as a limitation. Our fast and slow states are two memory lanes whose decay is adaptive and whose geometry is curved.
Two works are relevant to later stages but are not used in the core method. GeoMerge shows that closed-form or rational substitutes for matrix exponentials, specifically Cayley transforms, make geometric averaging practical for parameter merging, although only offline. The benchmark in AcCoRD covers conversational drift, including "triggered" preferences with satisfaction as low as 0.09. We leave conversational agents out of scope and treat that setting as a later target once the recommender result holds.
Research Program
The first component is the model itself. Item embeddings live on the hyperboloid and are parameterized by their spatial coordinates, with the time coordinate computed as $\sqrt{1/c+\|x\|^2}$. This keeps them on the manifold without retractions, so the training optimizer can be ordinary Adam. The user-state recurrence is differentiable and is trained end to end with a sampled-softmax next-item loss. Curvature $c$ is a single learned scalar. We deliberately do not use per-step curvature prediction as SINCERE does, and we test whether its absence costs accuracy.
The second component is an efficiency analysis. We count operations and measure latency for one streaming update and for one top-$k$ query. Four state trackers are compared: our recurrence; the same centroid computed by iterative Fréchet-mean steps with exponential and logarithmic maps; HMamba-FULL's recurrent step; and a Euclidean exponential moving average of the same form. The expected result is that our update and the Euclidean update have identical cost up to one square root per readout, while the exponential-map variant and HMamba-FULL cost measurably more.
The third component is an analysis of the dispersion signal. We check whether $r_t$ tracks externally measured drift. On MovieLens and on the point-of-interest check-in data, genre and venue-category labels let us measure drift as the shift in a user's category distribution between adjacent windows. If $r_t$ and the fast–slow disagreement do not correlate with that measure, the gate has no principled basis and the adaptive-forgetting claim fails.
Baselines
The naive reference point is an undecayed average of the user's item embeddings, together with a last-item model. Both are trained on the same embeddings. They show how much any drift handling helps. A popularity ranker will also be reported as a floor.
The strongest existing method on the recommendation benchmarks is HMamba, in both its FULL and HALF variants. The paper reports 3–11% improvements in HR@10, NDCG@10 and MRR@10 over baselines on ML-1M and on the New York, California and Texas point-of-interest datasets. It also reports convergence 1.8× faster than Mamba4Rec and 18–22% less GPU memory than attention models. The file contains no absolute values, so there is no reported result on Lacuna for HMamba's absolute HR@10 or NDCG@10 on ML-1M. We will therefore rerun HMamba ourselves under our protocol rather than quoting numbers. Beating HMamba means matching HMamba-FULL's NDCG@10 within noise over five seeds while achieving lower per-update latency than HMamba-HALF. On the high-drift user stratum, it additionally means outperforming both variants.
On the dynamic interaction benchmarks (LastFM, Wikipedia, Reddit, MOOC), the strongest drift-aware curved-space competitors are SINCERE and DyGMoCE. The file has no reported number on Lacuna for SINCERE's link-prediction scores on these datasets. For DyGMoCE, the file reports rank-1 placement on almost all datasets and the 27.5%/52.6% efficiency gains, but no absolute accuracy figures. Both will be rerun from their released settings.
Two Euclidean sequence models stand for the strongest flat approach: Mamba4Rec and SASRec, which the HMamba paper names but for which the file gives no artifact. Each will be given NS-DPO-style exponential time discounting on its training loss, following Right Now, Wrong Then, with the discount tuned by grid search.
All methods share the same chronological splits, the same embedding dimension, and the same number of hyperparameter trials, 30 per method. Latency is measured on the same GPU at batch size 1 for streaming and 1024 for bulk queries.
Four ablations isolate the key ingredients. The Euclidean twin uses the identical recurrence and gate on flat embeddings, and a spherical twin uses L2-normalized embeddings. Together they isolate whether curvature matters. A single-timescale version isolates the fast/slow split. A fixed-$\gamma$ version, tuned by grid, isolates whether the dispersion gate earns its place. The exponential-map centroid isolates whether the closed-form update loses any accuracy, which it should not, since it computes a slightly different mean at much higher cost.
Why not just…?
Why not just use a Euclidean moving average with the same decay? This is the most serious threat, because our recurrence is formally a moving average in a space whose inner product has one sign flipped. The file offers indirect evidence for curvature. HMamba finds that even partial hyperbolic processing gains 3–11% over Euclidean sequence models, and H²SeqRec reports 31.61% NDCG@1 gains on sparse Amazon data. Neither result compares matched-cost recurrences, however. The Euclidean-twin ablation decides the question directly. If the twin matches our model on the high-drift stratum and the dispersion signal works equally well in flat space, we will report that curvature is not needed for drift tracking and reframe the contribution as the adaptive two-timescale gate.
Why not just use HMamba-HALF, which already cuts the hyperbolic cost? HMAMBA-HALF recovers speed by leaving hyperbolic space before scoring, and its authors report that the FULL variant is more accurate. It also keeps a stationary transition kernel. Our model scores intrinsically with a single inner product, so it is not forced to trade speed against fidelity at the decision layer. The latency benchmark and the high-drift stratum comparison test exactly this claim. If HALF matches us on both, our geometric argument is moot for these datasets.
Why not just add NS-DPO-style time discounting to a strong flat sequence model? Discounting is nearly free and demonstrably helps. NS-DPO keeps reward accuracy above 60–70% after preference flips, whereas standard DPO drops to about 50–55%. Its known weakness is the global, hand-tuned discount. The spliced-flip experiment below compares a grid-tuned global discount against our per-user gate. If a single global discount recovers from flips as fast as the gate does, the gate is unnecessary.
Proposed Experiments
We will use chronological train/validation/test splits on ML-1M, the three point-of-interest datasets, LastFM, Wikipedia, Reddit and MOOC. Metrics are HR@10, NDCG@10 and MRR@10, all reported overall and on the top-quintile drift stratum, which is defined by the category-shift measure above. Efficiency is reported as update latency, query latency, and peak memory.
A controlled drift test follows the change-point design of NS-DPO. We splice the history of one user onto another's at a known point, creating a synthetic preference reversal. For each method we measure recovery time: the number of interactions until NDCG@10 on the new user's items returns to 90% of its steady-state value. A second version blends the two users gradually rather than switching abruptly.
The proposal is judged by three checks. First, per-update latency should match the Euclidean twin within about 10%, while being clearly below both HMamba variants and the exponential-map centroid. Second, the hyperbolic model should beat its Euclidean twin on the high-drift stratum and on recovery time. Third, $r_t$ should correlate with measured category drift, and the adaptive gate should shorten recovery time relative to the best fixed $\gamma$ without hurting stable users. If the first check holds but the second fails, the result is still useful as a negative finding: fast hyperbolic tracking is possible but buys nothing for drift. We would report it as such.
References
- Hyperbolic Neural Networks
- Unsupervised Hyperbolic Representation Learning via Message Passing Auto-Encoders
- Riemannian Continuous Normalizing Flows
- HMamba: Hyperbolic Mamba for Sequential Recommendation
- SINCERE: Sequential Interaction Networks representation learning on Co-Evolving RiEmannian manifolds
- Fast Mixture of Curvature-Aware Experts for Diverse and Dynamic Graph Topologies
- Hyperbolic Hypergraphs for Sequential Recommendation
- THINK: Temporal Hypergraph Hyperbolic Network
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization
- A Geometric View of Model Merging: Quotient Fréchet Averages from Toy Models to LoRA
- AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics