Minimum-Variance Combination of Time Scales for Per-User Preference Drift

Recommender systems have to decide how much a user's old behavior should still count. Most current answers use one global setting: a discount factor, a recency window, or a fixed split into "last few items" versus "everything." Users do not drift at one global speed. One person's taste is stable for years, while another's turns over every week. This note proposes a fix taken from microphone arrays. Summarize each user's history at several time scales, as running averages with different memory lengths. Shift each average forward to correct the lag its smoothing introduces. Then combine the averages with per-user weights computed in closed form, so the combined estimate has the smallest prediction error that user's recent behavior allows. In array processing, this is delay-and-sum beamforming made adaptive in the minimum-variance distortionless response (MVDR) style. In MVDR, sensor signals are time-aligned and then weighted so that the target passes through unchanged while everything else is suppressed as much as possible.

The two-sentence version is this. Each time scale is treated as a microphone that hears the user's current preference late and noisily. We undo each one's delay and then mix them with weights that each user's own error statistics determine. If the idea works, the drift rate stops being a hyperparameter tuned once for everyone and becomes a quantity estimated per user, cheaply and interpretably.

Core Claim

Non-stationary preference methods rely on a single exponential discount. The clearest example is Right Now, Wrong Then, which weights each preference pair by $\gamma^{T-t_i-1}$ inside a time-dependent Bradley-Terry model. Its authors note that the best $\gamma$ depends on a drift budget that is unknown and varies across users. Multi-scale recommenders such as CLSR and HierTCN hard-code the boundaries between scales: the last $k$ items, or session timeouts. They then fuse the scales with learned gates or pooling.

Our claim is narrower and testable. Under gradual drift, a long-memory average is not just noisier than a short one. It is systematically behind, by roughly its time constant. Two operations are each necessary:

  • Correcting that lag, which is the "delay" step.
  • Choosing per-user mixing weights by minimum variance, which is the "sum" step.

With both in place, a handful of fixed-scale channels should match or beat a tuned global discount and learned gating. The largest gains should come from users whose drift rate is far from the population average.

Concretely, let $e_i \in \mathbb{R}^d$ be the embedding of the item a user touched at continuous time $t_i$. Channel $k$ is an exponential moving average $m_k(t)$ with time constant $\theta_k$, for example one hour, one day, one week and one year. If the user's latent preference $p(t)$ moves with velocity $v$, then $m_k(t) \approx p(t) - \theta_k v$. This lag follows from the fact that an exponential kernel with time constant $\theta_k$ has mean age $\theta_k$. The lag-compensated channel is $\tilde m_k = m_k + \theta_k \hat v$, where $\hat v$ is estimated from differences between adjacent channels, much as double-exponential smoothing estimates trend. The user state is $u = \sum_k w_k \tilde m_k$, with

$$ w = \frac{(R+\delta I)^{-1}\mathbf{1}}{\mathbf{1}^\top (R+\delta I)^{-1}\mathbf{1}}. $$

Here $R$ is the $K\times K$ matrix whose $(j,k)$ entry is the average inner product of channel residuals $e_{\text{next}}-\tilde m_j$ and $e_{\text{next}}-\tilde m_k$ over that user's recent events. The vector $\mathbf{1}$ is all ones, which is the unit-gain constraint: the weights sum to one, so a preference held steadily across scales passes through undistorted. The term $\delta I$ is diagonal loading, the standard robustness fix in adaptive beamforming. It keeps the solution stable when a user has few events.

"Nulling" here happens across scales, not across directions in embedding space. A promotion burst inflates the residuals of the short channels, and the minimum-variance solution automatically down-weights them for that user.

Where the Pieces Come From

Array processing supplies the method. Directionally constrained minimization of power for speech signals generalizes a constrained minimum-power beamformer to broadband signals using FIR filter constraints and sample-matrix inversion. We reuse exactly this constrained-minimum-variance solution with sample-covariance estimation, applied to time-scale channels instead of sensors.

Two acoustic results also shape the design. RoboVox finds that unsteered averaging of robot microphones barely helps: 15.06% equal error rate versus 15.63–18.22% for single channels. That is the analogue of uniformly averaging time scales, and it is why we treat uniform averaging as the naive reference. Spille et al. report that MVDR can cancel the target when target and interferer share an angle. This rules out a tempting extension: actively nulling "abandoned" directions in embedding space would also erase a user who later returns to an old interest. We therefore constrain only the scale weights. We also evaluate a "return to old interest" slice explicitly, to catch any such cancellation.

On the recommendation side, several papers motivate the gap. CLSR and iFusion show that separating stable and transient interest helps, and that short-term noise leaks into long-term state. iFusion's low-pass/high-pass filters are a frequency-domain cousin of our channels. However, iFusion fuses through an iterative diffusion denoiser rather than a closed-form per-user estimator. LSAT shows that keeping separate long- and short-term components and ensembling their outputs beats merging them in parameter space. That supports combining at the representation or output level, which is what we do. FuXi-Linear makes timestamps a separate channel for long histories. It is complementary: our user state can feed such a backbone, and our channels are O(1) to update per event.

Research Program

We build the module in three forms of increasing coupling. First, a frozen-embedding version: item embeddings come from a trained SASRec-style model, and only the user state is replaced. This isolates the estimator. Second, an end-to-end version: gradients flow through the closed-form weights, which are differentiable through a small $K\times K$ solve, so item embeddings adapt to the combiner. Third, a hybrid in which $u$ and the individual $\tilde m_k$ are prepended as tokens to a time-aware transformer.

For each version we log the learned weights per user. That lets us check whether the implied effective memory length correlates with an independently measured drift rate. We measure drift rate as the displacement of a user's mean item embedding between consecutive months.

We also build a semi-synthetic simulator, because real logs have no ground-truth drift. Latent user preferences follow trajectories with per-user drift speeds drawn from a heavy-tailed distribution. Some users also receive promotion bursts, and some return to old interests. Items are sampled from these preferences using item embeddings learned on Amazon Beauty. This gives known true drift rates against which to check the recovered weights.

As a transfer test of the claim that the drift rate should be estimated rather than set, we apply the same estimator to NS-DPO's drifting preference datasets. There, the "channels" are reward-model estimates under several discount factors, and the minimum-variance weights replace the single $\gamma$.

Baselines

The naive reference point is uniform averaging of the same lag-corrected channels, which is the RoboVox-style unsteered sum. It shares every input with our method and differs only in the weights. Beside it we run a single exponential moving average whose $\gamma$ is tuned globally on validation data. This is the recommendation analogue of NS-DPO's discount and the cleanest statement of the hypothesis that one global drift rate is enough. We also run a sliding window of the most recent $k$ items, with $k$ tuned.

The strongest learned competitors are a time-aware flat transformer (SASRec with continuous time embeddings), CLSR's contrastive long/short dual encoder with attention fusion, and DCRec, the diffusion recommender. The file reports DCRec's gains over prior baselines only as relative numbers on Beauty and Toys: +3.45% HR@5 on Beauty, and +17.35% HR@5 and +17.87% NDCG@10 on Toys. No absolute HR or NDCG for DCRec was found on Lacuna, so we will re-run it under our splits. Likewise, CLSR reports only relative gains (GAUC over +0.01, NDCG over +4%, over +10.7% NDCG on Taobao). No absolute result for CLSR is found on Lacuna, and it too will be re-run.

For chronological incremental evaluation on MovieLens-1M, the strongest number in the file is LSAT's output ensemble at average AUC 0.7720, against 0.7650 for full retraining and 0.7594 for fine-tuning (LSAT). That result uses a LLaMA-7B TALLRec backbone. We will reproduce its 20-period protocol with our module attached to a small backbone. Beating it means matching 0.7720 at a fraction of the inference cost. A clear win would be exceeding it.

For fairness, every method uses the same chronological splits, the same item-embedding dimension, the same tuning budget of 30 trials per method on a held-out later time slice, and three seeds. Learned baselines get access to the same timestamp features.

Three ablations isolate the key ingredients. The first removes lag compensation, which reduces the method to classical minimum-variance forecast combination. The second replaces per-user weights with weights pooled across all users. The third removes diagonal loading. Beating the baselines means a statistically significant NDCG@10 improvement over the best tuned learned baseline overall, and a larger improvement in the top and bottom drift-rate quintiles. Our method should not lose in the middle quintile, where a global $\gamma$ is already near-optimal.

Why not just…?

Why not just tune one exponential discount, as NS-DPO does? The NS-DPO authors themselves state that the right $\gamma$ depends on a drift budget that is unknown and varies across users. A single value is a compromise that is too fast for stable users and too slow for volatile ones. Our quintile-stratified evaluation tests this directly. If the tuned global decay matches us in the extreme quintiles, the per-user premise is false and we will report that.

Why not just let a time-aware transformer or a learned gate figure out the weights? This is the real threat. A gate fed the same per-user residual statistics could in principle learn the minimum-variance rule. CLSR's evidence that disentangling scales beats entangled single sequences shows that structure helps, but it does not show that closed-form weights beat learned ones. We settle this by giving a learned gate exactly the residual covariance features our estimator uses. If the gate matches us everywhere, the contribution shrinks to an inductive bias. In that case we expect the gap to concentrate in users with short histories, and we will measure gain as a function of history length.

Why not just ensemble a long-term and a short-term model? LSAT shows that output ensembling works (0.7720 AUC on ML-1M) but doubles inference cost. Its fixed two-way split cannot adapt the mix per user. Our method is a per-user ensemble over several scales, costing one small linear solve. The LSAT-protocol comparison measures both accuracy and serving cost.

Why not just use plain minimum-variance forecast combination, with no delay step? This is the no-lag-compensation ablation. Our argument predicts that it loses specifically under gradual drift, where long channels are biased rather than merely noisy, and ties under stationarity. If it ties under drift too, the beamforming "delay" half adds nothing.

Proposed Experiments

Experiments run on Amazon Beauty and Toys and Yelp under the RecBole preprocessing described in DCRec, but with strictly chronological train, validation and test cuts so that drift exists at test time. They also run on MovieLens-1M under LSAT's 20-period protocol, and on KuaiRand-27K for long histories. Metrics are HR@10, NDCG@10 and AUC. These are reported overall, by drift-rate quintile, by history length, and on a "return to an interest dormant for more than 90 days" slice designed to catch target cancellation.

On the simulator, the primary check is whether the effective memory length implied by the learned weights correlates with the true per-user drift speed. We would consider a Spearman correlation above 0.5 meaningful. We also check whether promotion bursts are suppressed without hurting returning users.

On the NS-DPO datasets (TV-HH, UltraFeedback with preference flips, 2C NSGO), the reference points are those reported in Right Now, Wrong Then. DPO reward accuracy collapses to about 50–55% after a flip, while NS-DPO holds 60–70%; on AlpacaEval 2.0, NS-DPO reaches a length-controlled win rate of 8.93 versus 7.29. The question is whether minimum-variance weighting over several discounts matches NS-DPO's best $\gamma$ without tuning it.

Three outcomes would falsify the proposal. The first is a tuned global decay tying us in the extreme drift quintiles. The second is the no-lag ablation tying the full method under simulated gradual drift. The third is a feature-matched learned gate tying us at every history length. Any one of these would mean the beamforming structure adds nothing beyond what the alternatives already provide.