A deployed recommender goes stale. User tastes shift, new items arrive, and the meaning of old items changes, so a model trained last month slowly loses accuracy until someone pays for a full retrain. Practitioners usually face a blunt choice. They can retrain often, which is expensive, or retrain rarely, which costs accuracy. This note proposes a middle path. When a sequential recommender is retrained on successive time windows, the changes in its weights between retrains probably lie in a small number of recurring directions. If so, we can estimate those directions once, from a history of past retrains. Between retrains, we then run a cheap Bayesian filter over a handful of coordinates along them. A Bayesian filter is a running estimate of a quantity, plus how uncertain we are about it, that is updated as each new batch of interactions arrives.
The filter's uncertainty serves a second purpose: it signals when the cheap update no longer suffices and a full retrain is worth its cost. This turns the efficiency-versus-accuracy question into a measurable curve rather than a scheduling guess. The approach also suits sparse interaction data. A model with only eight to thirty-two free coordinates can be updated reliably from the thin trickle of interactions that arrives in a day.
Core Claim
We make two linked claims. The first is empirical. For a deep sequential recommender retrained on consecutive time periods, the differences between consecutive weight vectors are concentrated in a low-dimensional subspace, and that subspace stays useful for future periods. The second is methodological. Bayesian filtering in that subspace recovers most of the accuracy of periodic full retraining at a small fraction of its compute. Its two ingredients, per-direction drift rates and posterior variance, do better than a point-estimate update with a hand-tuned forgetting rate. Concretely, write the deployed weights at time $t$ as
$$w_t = \bar{w} + P z_t, \qquad z_t = z_{t-1} + \epsilon_t, \quad \epsilon_t \sim \mathcal{N}(0, \mathrm{diag}(\alpha)^{-1}).$$
Here $\bar{w}$ is the last fully retrained weight vector. $P$ is a $d \times k$ matrix whose columns are the top principal directions of historical retrain deltas $w^{(j+1)} - w^{(j)}$. $z_t \in \mathbb{R}^k$ are the drift coordinates we filter. $\alpha$ is a vector of per-direction precisions, meaning how fast each direction is allowed to drift, with Gamma priors on its entries.
How This Relates to Prior Work
The weight-space side borrows its machinery from Subspace Inference for Bayesian Deep Learning. That work builds a 5–20 dimensional subspace by applying principal component analysis (PCA) to one stochastic-gradient training trajectory, then samples a static posterior inside it. It reports 81.55% accuracy on CIFAR-100. We reuse the idea that a small subspace can carry useful posterior mass. However, we build the subspace from a different object: the sequence of deltas produced by retraining across time. We then use it for tracking over time rather than for averaging a fixed posterior. That paper also reports needing an ad hoc temperature to keep the posterior from collapsing. In our setting, the analogous knob is the drift precision $\alpha$, which we learn rather than hand-set.
The drift model follows Dynamic Bayesian Logistic Matrix Factorization (DBLMF). DBLMF places Gamma-prior Gaussian diffusion on each latent dimension of a linear matrix factorization, so that each dimension drifts at its own rate, and it scales variational inference by sampling negative interactions. DBLMF is the closest existing work. It applies drift to the user and item factors of a bilinear model and redoes inference across the full factorization at every time step. We apply the same prior to the coordinates of a deep sequence model's weight subspace. This keeps the nonlinear sequence encoder and makes each update cost $O(k^3)$, independent of the number of users.
Linear Response Methods for Accurate Covariance Estimates from Mean Field Variational Bayes supplies a cheap way to repair the variance that diagonal, mean-field approximations underestimate. Its main cost, an $O(P^3)$ matrix inversion, is harmless when $P = k \le 32$. Its known failure, biased mean-field means under strong dependencies, is one we can check directly against sampling in the same small subspace.
A Shared Low-Rank Adaptation Approach to Personalized RLHF motivates a hierarchical extension. That paper shares one low-rank projection across users, gives each user a small personal component, and proves that sharing cuts per-user sample complexity under static preferences. We add optional per-user offsets in the same drift subspace, shrunk toward the global coordinates, so that sparse users borrow strength while dense users can move.
Two papers constrain the design. Modeling Sequences as Distributions with Uncertainty for Sequential Recommendation (DT4SR) shows that representing preferences as distributions helps sparse items. It reports a 32.4% relative Recall@1 gain on Amazon Toys. Its uncertainty, however, comes from network outputs, not from a posterior over a model that is going stale. Oracle-guided Dynamic User Preference Modeling for Sequential Recommendation (Oracle4Rec) handles drift inside the training objective by using future sequences as a training signal, and stays a point model. Neither addresses what happens after deployment.
We use Multi-Behavioral Sequential Recommendation (ConvRec) as a base architecture. ConvRec replaces quadratic-cost attention with pyramidal strided convolutions that run in time linear in sequence length and fuses interaction timestamps. Linear cost matters because one of our baselines simply feeds the model much longer histories.
Research Program
The first stage is a measurement study, and it can falsify the premise. We take ConvRec, and a SASRec-style attention model as a second base, and simulate a production schedule. Each model is trained on the first block of time and then retrained from a warm start at every period boundary, for example monthly on MovieLens-20M and quarterly on the Amazon categories. We record the weight deltas. We then measure three things. The first is how much each frozen model's accuracy decays over subsequent periods. The second is how much of the variance of the deltas the top $k$ principal directions explain. The third, which is what matters, is how much of a future retrain's accuracy gain is recovered by projecting that retrain's delta onto a subspace estimated from earlier deltas.
If frozen models do not decay, or future deltas do not fall in past subspaces, the rest of the program is unnecessary, and we would report that as the finding.
The second stage builds the filter. Recommendation likelihoods, whether a softmax over items or a pairwise ranking loss, are not Gaussian. We therefore update $z_t$ with a Laplace approximation per batch, in the style of an extended Kalman filter: take a few Newton steps on the batch log-likelihood plus the Gaussian prior propagated from $t-1$, then use the local curvature as the new precision. Per-batch gradients in $z$-space cost one backward pass, projected through $P$. We place drift precisions $\alpha$ under Gamma priors and update them by variational expectation steps, as in DBLMF.
We compare three inference methods inside the same subspace, which together span the efficiency-accuracy dial at small scale. The cheapest is a diagonal mean-field update. The middle option is that same update with a linear-response covariance correction. The most expensive reference is elliptical slice sampling, a simple Markov chain Monte Carlo method that needs no step-size tuning, run as in Izmailov et al.
The hierarchical variant adds per-user coordinates $z_{u,t}$ with prior $\mathcal{N}(z_t, \tau^{-1} I)$. Here $\tau$ is the precision that controls how strongly each user is pulled toward the global drift. It is filtered only for users with new interactions.
The third stage turns uncertainty into a retrain decision. We use two signals. One is the posterior predictive entropy change. The other is the out-of-subspace gradient fraction: the share of the batch gradient norm orthogonal to the columns of $P$, which indicates drift the subspace cannot express. When either crosses a threshold, we trigger a full retrain and re-estimate $P$ with the new delta. Sweeping the threshold traces a curve of accuracy against cumulative compute, which is the object we want to improve.
Baselines
The do-nothing reference is the frozen model: trained once, never updated, evaluated over every later period. It anchors the decay curve, and any method must close a meaningful fraction of the gap between it and the next baseline.
The accuracy ceiling at high cost is periodic full retraining at several frequencies, from once per evaluation horizon to every period. Each run uses the same warm-start recipe and tuning budget as our base model, and each is plotted as one point on the accuracy-versus-compute plane. This is the strongest competitor in practice, because it is what deployed systems do. Beating it means sitting strictly above its frontier: matching its NDCG@10 at lower total GPU time, or exceeding it at equal time.
For reference to published strength, we report the architectures' published numbers. These come from leave-one-out offline splits, not deployment splits, so they are not directly comparable and serve only to confirm that our reimplementations are competitive. Oracle4Rec reports HR@1 of 0.2257 on MovieLens-100K against DualRec's 0.2097. ConvRec reports up to 4.08% HR@10 and 7.82% NDCG@10 gains over ProxyRCA on Amazon Beauty, and 6,096 MB memory against 14,280 MB at sequence length 500 on Amazon Games. DBLMF reports outperforming eALS by over 139% in NDCG@5 on its MovieLens-20M subset. No reported result was found on Lacuna for any of these methods under a rolling temporal deployment protocol, so we will run them ourselves on the same splits.
Three fine-tuning baselines receive the same per-period compute as our filter, measured in matched gradient evaluations. The first is low-rank adaptation (LoRA) on each period's data. The second is full-parameter fine-tuning for a few steps on recent data. The third is DBLMF itself, run as a dynamic Bayesian factorization baseline that already combines drift, sparsity, and scalable inference without a deep encoder.
Two ablations isolate our ingredients. In the first, a point-estimate ablation, we optimize $z_t$ by stochastic gradient descent with exponential forgetting and a tuned learning rate. Its tuning budget equals our filter's hyperparameter search. This removes the posterior and the learned per-direction drift rates while keeping the subspace. In the second, a subspace ablation, we replace $P$ with a random orthonormal basis or with the PCA of a single static training trajectory. This tests whether drift directions specifically matter. Beating the point ablation by a reliable margin on sparse users and abrupt-drift periods is the result that justifies the Bayesian machinery. Merely matching it would not.
Why not just…?
Why not just retrain more often? For a large enough budget, frequent retraining will win on accuracy, and we do not claim otherwise. The claim concerns the frontier: below that budget, a filtered subspace should dominate. The retrain-frequency sweep in our baselines settles this directly. If the retraining frontier lies on or above ours everywhere, the proposal fails.
Why not just let the sequence model handle drift through its input history? This is a real threat. A model that reads a user's recent items already adapts to that user's recent behavior without changing weights. Our premise is that the shared weights themselves go stale, for example in how they map items to intents as catalogs and populations change. The frozen-model decay measurement in stage one tests this. We also add a long-history variant using ConvRec's linear-time encoder at sequence length 500. If longer histories close the gap as well as filtering, we will say so.
Why not just fine-tune with LoRA or SGD on recent data? Low-rank fine-tuning already restricts updates to a small space. On sparse daily data, though, it has no principled step size, no memory of how fast each direction historically moved, and no signal for when it is out of its depth. The point-estimate ablation and the matched-compute LoRA baseline measure whether those properties pay off. We expect the gap to be largest for users with one or two interactions per period and in periods with abrupt shifts.
Why not just use distributional embeddings like DT4SR or a deep ensemble for uncertainty? DT4SR's variance describes breadth of preference under a fixed model. It does not say whether the model itself has drifted. An ensemble of $M$ members multiplies both training and update cost by $M$. We evaluate both as retrain triggers, replacing our two signals with their predictive variance, at matched compute. The test is whether their uncertainty predicts next-period accuracy loss as well as ours does.
Proposed Experiments
We use four datasets. MovieLens-20M spans years and was used temporally by DBLMF. Amazon Beauty, Toys, and Games provide sparse, cold-start-heavy categories with long histories. All are split by absolute time into a pre-deployment block, a block of historical retrains used to estimate $P$, and a later evaluation block that the subspace never saw. We report NDCG@10 and HR@10 per evaluation period, cumulative GPU-seconds, and accuracy stratified by user activity, splitting out users with at most two interactions in the period.
We run three decisive checks. The first is whether projecting held-out future retrain deltas onto the past-estimated $P$ recovers most of their accuracy gain at $k \le 32$; this is the premise. The second is whether the filter's accuracy-compute curve lies above the full-retraining frontier, with the point-estimate ablation falling measurably below the filter for sparse users. The third is whether the out-of-subspace gradient fraction and posterior entropy predict the next period's accuracy loss, measured by rank correlation across periods. If they do, a schedule triggered by these signals should beat any fixed retraining schedule at equal total compute.
A secondary analysis compares the mean-field update, the linear-response correction, and slice sampling within the subspace. It reports whether linear response recovers the sampling covariance in this regime, or whether it fails the way Giordano et al. warn it can when mean-field means are biased.
References
- Subspace Inference for Bayesian Deep Learning
- Dynamic Bayesian Logistic Matrix Factorization for Recommendation with Implicit Feedback
- Linear Response Methods for Accurate Covariance Estimates from Mean Field Variational Bayes
- A Shared Low-Rank Adaptation Approach to Personalized RLHF
- Modeling Sequences as Distributions with Uncertainty for Sequential Recommendation
- Oracle-guided Dynamic User Preference Modeling for Sequential Recommendation
- Multi-Behavioral Sequential Recommendation