When a restaurant is full, a hospital ward has no beds, or a bus drives past a crowded stop, people take their second choice. A system that learns preferences from logged choices then sees what looks like a change of heart. Much of the current work on tracking changing preferences, for example down-weighting old feedback so the model follows the most recent behavior, implicitly assumes every option was always available. That is the "infinite capacity" assumption that operations research has long made for mathematical convenience. Under that assumption, a surge of substitutions during a busy period is read as preference drift. The model then chases the substitute, and it recommends the wrong thing once capacity frees up. The reverse error also occurs: models that respect capacity treat preferences as frozen, so they cannot follow users whose tastes really do change.
This note proposes a single model that learns time-varying user preferences from choices that were shaped by finite, shared capacities, in the realistic case where we never observe which option was unavailable to which person. The core claim is that capacity-induced shifts and true drift leave different fingerprints. Capacity-induced shifts are item-level, synchronized across all users competing for the same saturated option, and appear only when that option's aggregate usage hits its limit. True drift is user-specific and persists after congestion clears. A model that forces its predicted assignments to respect capacities, while letting preferences evolve under a prior that favors rare changes, should separate the two. Doing so should also remove the need to hand-pick a forgetting rate.
Core Claim and Mechanism
The starting point is SiMCa: Sinkhorn Matrix Factorization with Capacity Constraints. SiMCa treats observed allocations, such as patients to hospitals, as the solution of a capacity-constrained assignment problem. It makes that assignment differentiable with entropic optimal transport (the Sinkhorn algorithm, which turns a hard matching into a smooth one by repeatedly rescaling rows and columns). This lets it recover latent preferences that capacity limits had hidden. SiMCa supplies our method, but it is a batch model with fixed user embeddings. We make it temporal.
At each period $t$, let $u_i(t)$ be user $i$'s preference vector, $v_j$ item $j$'s features, $d_{ij}$ a known cost such as distance, and $c_j(t)$ the item's capacity in that period, known or estimated from aggregate counts. The predicted soft assignment $A(t)$ is the entropic transport plan between users and items, including an "outside option" column that absorbs unserved demand. Its affinities are $u_i(t)^\top v_j - d_{ij}$, and its column totals are bounded by $c_j(t)$. We fit
$$\min_{u,v} \sum_t \ell\big(Y(t), A(t;u(t),v)\big) + \lambda \sum_i \sum_t \lVert u_i(t) - u_i(t-1)\rVert_2 ,$$
where $Y(t)$ holds the observed choices and $\ell$ is the negative log-likelihood of those choices under $A(t)$. The second term is a group penalty on period-to-period changes. It prefers each user's preferences to stay constant and change in a few discrete jumps, which is the "rare real drift" prior.
The capacity constraint explains synchronized substitutions for free. When item $j$ saturates, the transport plan itself pushes lower-affinity users to their next-best option, so no change in $u_i$ is needed to fit that behavior. The drift penalty then refuses to spend on changes that the constraint already explains. Because $\lambda$ trades off total variation against fit, it can be tuned on held-out likelihood instead of requiring prior knowledge of how much drift exists. Right Now, Wrong Then names this "variation budget" problem as an open weakness of exponential recency weighting.
The identification argument has a clear failure mode, and we will test it rather than hide it. If a real preference change happens to all users at once, in the same direction, and exactly when an item saturates, the two explanations are confounded.
Relation to Prior Work
Two papers relax infinite capacity in recommendation, and both hold preferences fixed. Recommendation under Capacity Constraints adds a differentiable penalty on expected item usage to matrix factorization. It shows this beats post-hoc truncation, and it explicitly lists dynamic propensities and fluctuating inventories as future work. SiMCa solves the inverse problem of recovering preferences from bottlenecked allocations, again on static data. Our proposal is the temporal version of the inverse problem. This is a different question from the forward one of recommending under capacity.
On the drift side, NS-DPO applies an exponential weight $\gamma^{T-t_i-1}$ to preference pairs. It shows that stationary training collapses to near-chance accuracy after preference flips. Its generation space is unconstrained, so any shift in behavior counts as evidence of drift. We use NS-DPO's recency-weighting idea as a baseline transplanted into the choice setting, not as a method to beat on LLM alignment. Towards Explainable Conversational Recommendation separates long-term from in-session preferences with its global and local views. Our persistent-versus-transient split is similar in spirit, but it is grounded in a physical constraint rather than session boundaries.
AcCoRD motivates the problem from the agent side. It shows frontier LLM agents fail when users want options inventory cannot satisfy, but it evaluates single users and does not model contention among many users. Multi-user contention is exactly the structure our identification relies on.
Transit work supplies the setting and a secondary hypothesis. Bus Frequency Optimization: When Waiting Time Matters in User Satisfaction assumes buses never fill and lists that as its main limitation. Capturing Travel Mode Adoption in Designing On-demand Multimodal Transit Systems models adoption as a static travel-time threshold. Together they suggest that capacity failures, such as being passed by a full bus, may cause real drift toward the outside option. Our model can represent this with an extension where $u_i(t)$'s affinity for the outside option depends on how often the user was unserved in earlier periods.
Research Program
The work has three parts. The first builds the temporal Sinkhorn model: a differentiable implementation with an outside-option column, capacities that change by period, and the group drift penalty. It is fit by gradient descent through unrolled Sinkhorn iterations, warm-started across periods so cost grows linearly in the number of periods.
The second builds semi-synthetic testbeds where ground truth is known. We take real timestamped check-ins from Foursquare and Gowalla, the datasets used by Christakopoulou et al., and fit time-varying ground-truth user embeddings. We then plant true drift jumps in a random subset of users. Next we impose capacities that bind only in chosen peak periods, and generate observed choices by capacity-constrained assignment. Finally we discard per-choice availability, so only choices and aggregate capacities are visible to learners. We extend SiMCa's synthetic hospital-allocation setup the same way. As a transit stretch goal, we use the Singapore bus records from Mo et al. with simulated vehicle capacities, where bypassed riders appear as later boardings on alternative routes.
The third adds the endogenous extension described above. We test whether the model can recover a planted rule of the form "users who were unserved $k$ times shift toward the outside option."
Baselines
The naive reference point is a static preference model fit on all data with no capacity awareness: plain matrix factorization over the whole log. It shows how badly both effects hurt when both are ignored.
The strongest competitor for tracking drift is a recency-weighted unconstrained choice model. This is matrix factorization with NS-DPO-style exponential weights on each observation, with $\gamma$ tuned on held-out periods. It is the natural way the drift literature would attack this data. The file reports no result for it on capacity-bound choice data; the closest reported figure is NS-DPO's own reward accuracy staying above 60–70% under preference flips, where standard DPO collapses to about 50–55% on non-stationary UltraFeedback (Right Now, Wrong Then). That number is from a different domain and serves only as context.
The strongest capacity-aware competitor is SiMCa refit on a sliding window, plus Cap-MF/Cap-GeoMF from Christakopoulou et al. For reference, Cap-GeoMF on Foursquare reduced capacity loss from 2.35 to 0.15 while RMSE rose from 0.66 to 0.97, and SiMCa reports assignment-recovery F1 near 1.0 on its static synthetic hospital data. We found no reported result on Lacuna for any method on time-varying, capacity-censored choice data, so every temporal number in this proposal will be new.
Two ablations isolate the key ingredients. One removes the capacity constraint but keeps the drift penalty. This tests whether the constraint is what prevents false drift. The other removes the drift penalty and fits each period independently. This tests whether sharing strength across periods matters.
An oracle upper bound is a conditional logit model given the true per-choice availability sets.
All methods get the same data splits, the same embedding dimension, and the same hyperparameter search budget of equal trials on the same validation periods. Beating the field means lower error in recovering true preference trajectories and fewer false drift detections than both the recency-weighted and windowed-SiMCa baselines, without losing recall on planted true drift.
Why not just…?
Why not just recency-weight an unconstrained model and rerank for capacity afterwards? Reranking fixes what is recommended, but not what is learned. Substitutions during peaks are still absorbed into the preference estimates as drift. Christakopoulou et al. already found that joint training beats post-hoc truncation in the static case. Our factorial experiment tests the temporal version directly. If the reranked recency model matches us on off-peak accuracy after a congestion episode, our central claim fails.
Why not just condition the choice model on which options were available? This is the textbook fix and a real threat whenever availability is logged per choice. That is why we include it as an oracle. Our target is the common case where availability for each person is unobserved: who was turned away at a full venue, which rider a full bus skipped. There, only aggregate capacity is known. The informative measurement is how much of the gap between the unconstrained baseline and the oracle our model closes. If it closes almost none, the idea is not worth pursuing.
Why not just refit a static capacity model on a sliding window? This will likely work reasonably well when drift is slow. We expect it to lose in two ways. Short windows make per-period capacity inference noisy. Fixed window lengths also blur abrupt jumps, the same window-length problem that Right Now, Wrong Then documents for discounting. Sweeping drift speed and jump size will show where, if anywhere, windowing matches the drift prior. We will report the regime where it matches.
Decisive Experiments
The central test is a two-by-two design crossing whether capacity binds with whether true drift occurs. The model should report near-zero drift in the cell where capacity binds and no true drift occurs, where the recency baseline should report false drift. It should match the recency baseline in the cell where drift occurs and capacity does not bind.
Metrics are:
- correlation between recovered and true preference trajectories;
- precision and recall of detected drift events;
- top-k accuracy of recommendations in off-peak periods immediately after congestion clears, which is where confusing the two effects costs the most;
- capacity loss as defined by Christakopoulou et al.
A deliberate stress test plants drift that is synchronized across users and coincides with saturation. This is where we predict the method fails, and we will measure how gracefully. For the endogenous extension, success means recovering the planted threshold $k$ from choices alone.
References
- SiMCa: Sinkhorn Matrix Factorization with Capacity Constraints
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Recommendation under Capacity Constraints
- Towards Explainable Conversational Recommendation
- AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics
- Bus Frequency Optimization: When Waiting Time Matters in User Satisfaction
- Capturing Travel Mode Adoption in Designing On-Demand Multimodal Transit Systems