Recommenders and assistants are retrained on fresh logs because people's tastes change. Each retraining round quietly asks a data-selection question: which of the old and new interactions should the model learn from next? Current practice either feeds in the whole new block of data or down-weights old data by age. We propose answering the question geometrically. First, measure how the population of interactions moved between the last two time periods, using optimal transport, which is the cheapest way of moving one cloud of points onto another. Then extrapolate that motion one step forward to get a forecast of the next period's data. Finally, pick a small training subset, drawn from both old and new data, whose distribution matches the forecast rather than the present.
The claim is simple. A subset chosen to match a forecast of the next period should generalize to that period better than three alternatives at the same budget: a subset matched to the current period, a subset matched to an age-discounted mixture of past periods, or the full uncurated block. The advantage should be largest when drift is gradual and directional, and should vanish under abrupt preference flips. That makes the claim easy to falsify.
Why this gap is real
Two literatures are currently disjoint. Optimal transport (OT) data selection chooses a training subset whose distribution is close, in Wasserstein distance, to a target set. TAROT does this with whitened gradient features and beats additive influence scores on multimodal targets. A Critical Look at Targeted Instruction Selection shows that most targeted selection methods approximately minimize this Wasserstein distance, and that unbalanced OT overtakes greedy selection as budgets grow. Closest to our setting, GORACS selects coresets for LLM-based recommenders by bounding test loss with an OT term plus gradient norms. It reports gains of about 5–7% in NDCG@5/HR@5 over DEALRec at 15% of the training FLOPs, and finds that small coresets beat full-data training by reducing popularity bias. All three treat the target distribution as fixed, and GORACS explicitly leaves temporal dynamics to future work. These papers supply our selection machinery and motivate the claim that curation beats bulk ingestion.
The drift literature, in turn, never selects data. Right Now, Wrong Then (NS-DPO) weights preference pairs by an exponential factor $\gamma^{T-t_i-1}$ of their timestamp. By the authors' own account, this cannot tell a high-quality old point from a noisy new one. PESO trains a single evolving low-rank adapter with a KL anchor across chronological stages, and CSTRec keeps fixed historical and current interest pools. Both ingest each incoming block whole. These papers define the regime and the strongest training backbone. Walk2Pers models individual user drift as a directed geometric walk with memory rather than a memoryless jump. It supplies the intuition we lift to the population level: drift has a direction, so it can be extrapolated.
The obvious combination is to run GORACS on each stage with a time-discounted target. We treat that as our main competitor, not our contribution. The new ingredient is the forecast target.
Core Claim and Mechanism
At stage $t$, embed every training example in a feature space. We use random projections of the gradient of the recommender's adapter parameters, following the finding in A Critical Look at Targeted Instruction Selection that gradient features track test loss closely. Let $\mu_{t-1}$ and $\mu_t$ be the empirical feature distributions of the last two stages. We solve an entropy-regularized OT problem between them with Sinkhorn iterations, as in Guzman-Nateras et al., and obtain a coupling $\pi$. For each current point $y$, the coupling gives a backward displacement: where $y$'s mass came from. The forecast distribution moves each point further along that path:
$$\hat{\mu}_{t+1} = \text{law of } \; y + \alpha\,\big(y - \mathbb{E}_{\pi}[x \mid y]\big), \quad y \sim \mu_t,$$
Here $x$ ranges over points of $\mu_{t-1}$, $\mathbb{E}_{\pi}[x \mid y]$ is the barycentric average of the old points coupled to $y$, and $\alpha \ge 0$ is a step size. Setting $\alpha = 0$ recovers "match the present." We choose $\alpha$ by backtesting: forecast stage $t$ from stages $t-2$ and $t-1$, and pick the $\alpha$ that minimizes the Wasserstein error. This requires no model training.
We then select a subset $S$ of fixed budget $k$ from the pool of all retained history plus the new block, minimizing the unbalanced OT cost $\mathrm{UOT}(\nu_S, \hat{\mu}_{t+1})$, where $\nu_S$ is the empirical distribution of $S$. The GORACS gradient-norm term and its initialize-then-refine swapping algorithm handle scale. Because the pool contains old data, forecast mass that lands in regions populated by past interactions pulls those examples back in. This is how returning interests and dormant users get represented without a hand-set history/current split. The selected subset trains the PESO adapter in place of the full block.
Research Program
The work has three parts. First, a diagnostic study: on each benchmark, how well does $\hat{\mu}_{t+1}$ predict the true next-stage distribution, compared with $\mu_t$ and with age-discounted mixtures of past stages? This settles cheaply whether population-level drift is directional enough to extrapolate. Second, the selection-and-training study described below. Third, a stress test on synthetic and preference-flip data, to map when extrapolation helps (gradual, directional drift) and when it hurts (sudden reversals). There we test a guard that shrinks $\alpha$ toward zero when the backtest error exceeds the error of $\mu_t$.
Baselines
The naive reference is PESO trained on each full incoming block with no selection (PESO). This is what current continual recommenders do. PESO reports average gains of 3.7% to 6.3% on Hit@5/10 and NDCG@5/10 over the best competing continual methods. No reported result was found on Lacuna for PESO's absolute metric values on these splits, so we will reproduce it. Beating it means equal or better next-stage Hit/NDCG at a fraction of the training FLOPs.
Uniform random subsets of the combined pool at the same budget $k$ test whether any curation matters. The Critical Look paper warns that random selection catches up at large budgets, so we sweep $k$ from roughly 1,024 examples (the GORACS regime) up to the full block. We report where the curves cross.
The strongest existing method is GORACS applied per stage, run in two variants. In the first, the target is the current block. In the second, the target is an age-discounted mixture of all past stages with weight $\gamma^{t-s}$ on stage $s$. The second is the "obvious" combination of OT selection and NS-DPO-style decay, with $\gamma$ tuned on the same backtest budget used for $\alpha$. GORACS reports about 5–7% NDCG@5/HR@5 over DEALRec in static settings. No reported result was found on Lacuna for GORACS in a streaming setting. Beating it means a statistically reliable next-stage improvement across seeds, on at least three of the four datasets, at identical $k$, identical features, identical solver, and identical adapter training.
A loss-reweighting baseline applies NS-DPO's exponential weights (Right Now, Wrong Then) to the recommender loss over all data, with no selection.
The key ablation replaces the forecast with $\alpha = 0$ while keeping the history pool. This isolates the extrapolation step from the effect of letting old data compete for the budget.
Why not just…?
Why not just down-weight old data by age, as NS-DPO does? Age weighting scores every example from the same period identically. It therefore cannot keep a valuable old example or drop a redundant recent one, and GORACS shows that uncurated data carries popularity bias that weighting does not remove. Still, this is a fair competitor on accuracy, not only on cost. The loss-reweighting baseline and the decayed-target GORACS baseline will show whether geometry adds anything beyond recency.
Why not just run GORACS on the newest block? This is the most serious threat. If drift between stages is small relative to within-stage noise, matching the present may be as good as matching a forecast. The diagnostic study measures this directly by comparing forecast error with the error of simply using $\mu_t$. The $\alpha = 0$ ablation then measures it in downstream accuracy. If forecasts do not beat $\mu_t$ in the diagnostic, we will report that and fall back to the simpler contribution of streaming OT selection with a history pool.
Why not just replay a random reservoir of old data alongside the new block? Random replay addresses forgetting but spends budget in proportion to how often items appear. That favors popular items and ignores where interests are heading. CSTRec's retention and acquisition split (RA/LA) lets us check separately whether our method retains dormant users as well as replay does while acquiring new interests faster.
Proposed Experiments
We follow PESO's chronological protocol: a 60% base stage followed by four 10% incremental stages. We run it on Amazon 2023 Musical Instruments, Movies & TV, and Books, on Yelp, and on MovieLens-1M. The LLM recommender backbone is a LLaMA-3.2-3B-class model, as used by GORACS. We report Hit@5/10 and NDCG@5/10 on the next stage, retention and acquisition following CSTRec, and total selection-plus-training FLOPs.
The decisive checks come in order. First, forecast Wasserstein error must beat both $\mu_t$ and the best decayed mixture on at least most stages. Second, forecast-matched selection must beat decayed-target GORACS at small and mid budgets. Third, the gain must correlate across stages with how much the forecast beat $\mu_t$. If the gain appears where forecasts are no better, the mechanism is something other than extrapolation.
As a transfer test, we select preference pairs on NS-DPO's gradual-drift and sudden-flip settings (TV-HH, UltraFeedback-LM). NS-DPO reports length-controlled win rate 8.93 versus DPO's 7.29 on UltraFeedback-LM. We expect forecast selection plus DPO to help under gradual drift. Under sudden flips we expect it to fail without the shrinkage guard, and to recover with it.
References
- TAROT: Targeted Data Selection via Optimal Transport
- A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn’t)
- GORACS: Group-level Optimal Transport-guided Coreset Selection for LLM-based Recommender Systems
- Right Now, Wrong Then: Non-Stationary Direct Preference Optimization under Preference Drift
- Continual Low-Rank Adapters for LLM-based Generative Recommender Systems
- Leveraging Historical and Current Interests for Continual Sequential Recommendation
- Beyond Markovian Drifts: Action-Biased Geometric Walks with Memory for Personalized Summarization
- Cross-Lingual Event Detection via Optimized Adversarial Training
- Capturing User Interests from Data Streams for Continual Sequential Recommendation