Drift Phenotypes: Letting Recommenders Learn How Fast Each Kind of User Changes

Every recommender that follows a user over time has to decide how much of the past to trust. If it trusts too much, it keeps recommending what the user liked last year, a failure called preference inertia. If it trusts too little, one gift purchase or one accidental click derails it. Current sequential recommenders make this choice either with one global forgetting rule or implicitly inside a single neural network shared by all users. Epidemiology has a well-tested tool for a closely related problem. Group-based trajectory modeling fits a small number of latent groups, each following its own curve over time, and assigns every person a probability of belonging to each group. It exists because a single population-average curve can hide subgroups whose trajectories point in opposite directions.

We propose to apply that tool to the rate at which a user's taste moves, rather than to the user's taste itself. We first summarize each user's history as a low-dimensional curve of "how much did my interests shift in this window." We then fit a growth mixture model to these curves, which yields a few drift phenotypes, for example users whose tastes settle, users who keep wandering, and users who return to old interests in cycles. Each user's group membership is updated online from their partial history, and the recommender uses it to set how quickly it forgets. The expected payoff is largest where individual evidence is weakest: users with short histories, and the first few interactions after a change, when the system has to decide whether the change is noise or a real new direction.

Core Claim

The claim is narrow and falsifiable. On standard sequential recommendation benchmarks, users differ in how fast they drift, and this variation is multimodal, not a single bell curve. If so, a mixture prior over drift trajectories will set per-user forgetting better than three alternatives: a global rule, an unpooled per-user estimate, and a unimodal shrinkage estimate. The advantage should show up mainly for short-history users and in two measurable behaviors: recovering quickly after a sustained taste shift, and staying put after a one-off interaction. If the drift distribution turns out to be unimodal, the mixture should add nothing over single-population shrinkage. That outcome would be a clean negative result, and it mirrors the original argument for trajectory mixtures over mixed models.

Relation to Prior Work

The motivation comes from the trajectory literature. Effect of Hemoglobin A1c Trajectories on Future Outcomes in a 10-Year Cohort With Type 2 Diabetes Mellitus shows that patients with identical multi-year mean HbA1c but unstable or rising trajectories have much worse outcomes. That is the "average hides the subgroups" argument, and we transfer it from blood sugar to drift speed. The method comes from two sources. Trajectories of Outpatient Service Utilisation of Hypertensive Patients in Tertiary Hospitals in China uses latent class growth analysis, in which members of a class share one exact curve. Trajectories of Depressive Symptoms Among Web-Based Health Risk Assessment Participants uses growth mixture modeling, which allows individual variation around each class curve, and finds that variation necessary. We adopt the growth-mixture form for the same reason. These works fit univariate clinical measurements offline. That constrains our design: rather than forcing a mixture model onto thousand-dimensional item embeddings, where expectation-maximization fitting is known to be brittle, we compress each history into a one- or two-dimensional drift signal, the setting where these models are known to work.

On the recommendation side, A survey on sequential recommendation describes self-attentive encoders such as SASRec, which read the ordered history with one network for all users. Preference Trajectory Modeling via Flow Matching for Sequential Recommendation (FlowRec) goes further. It learns a deterministic flow from a history-based prior to the next-preference embedding and outperforms diffusion recommenders such as DiffuRec. Diffusion-based alternatives include Preference Diffusion for Recommendation and From Agnostic to Specific: Latent Preference Diffusion for Multi-Behavior Sequential Recommendation. All of these model each user's path individually and contain no group structure over how paths evolve. We do not replace them; they serve as backbones, and our contribution is a module that controls how much history they see. Xiting Wang et al. identify preference inertia, meaning that long histories drown out recent shifts. This motivates the problem, but that work does not say how much to discount the past for whom.

The closest structural relative is Latent Preference Bandits, which infers a user's latent class online to cut regret. Its classes describe what users prefer, under stationary or piecewise-stationary assumptions. Ours describe how preferences move, with a growth curve inside each class. Papers titled "group recommendation," such as Deep Modeling of Group Preferences for Group-Based Recommendation and Attention-based Group Recommendation, aggregate the tastes of people deciding together and are unrelated despite the shared word. Evaluating Rule-based Programming and ReinforcementLearning for Personalising an Intelligent System shows that users' preferences form partly in response to what the system shows them. This is a constraint on interpretation: drift measured in offline logs mixes the user's own change with the logging policy's influence, so our phenotypes describe logged behavior, not intrinsic psychology.

Research Program

Drift signal. We freeze the item embeddings of a trained backbone and cut each user's history into windows of a fixed number of interactions. We use count-based windows because calendar gaps are irregular; a calendar-time variant is tested separately. For user $i$ and window $w$, the drift is $d_{iw} = 1 - \cos(m_{i,w}, m_{i,w-1})$, where $m_{i,w}$ is the mean embedding of items in window $w$. A second channel, the fraction of repeat or same-category interactions, captures cyclic return.

Mixture model. We fit a growth mixture model with class-specific quadratic curves and a random intercept per user:

$$d_{iw} = \beta_{0k} + \beta_{1k} w + \beta_{2k} w^2 + b_i + \epsilon_{iw}, \quad b_i \sim \mathcal{N}(0, \tau_k^2),\ \epsilon_{iw} \sim \mathcal{N}(0, \sigma_k^2).$$

Here $k$ indexes the $K$ classes, the $\beta$ terms are that class's intercept, slope, and curvature, $b_i$ is user $i$'s personal offset, and $\sigma_k$ is the within-class window-to-window noise. The number of classes $K$ is chosen by BIC together with stability across random restarts and across time-split halves of the user population. We report all criteria, because they often disagree.

Online use. At prediction time, user $i$'s class posterior $\pi_{ik}$ is computed from however many windows they have so far, using the standard Bayes update, so it sharpens as history accumulates. Each class carries a learned forgetting rate $\lambda_k$. The backbone reweights history item $j$ at age $a_j$ by $\exp(-\bar\lambda_i a_j)$, where $\bar\lambda_i = \sum_k \pi_{ik}\lambda_k$. For SASRec this weighting is applied as a bias on attention logits; for FlowRec it is applied to the history pooled into the flow's prior. The class noise level $\sigma_k$ gives a principled test for surprise. A new window whose drift sits far above the class-expected curve, relative to $\sigma_k$, is held in a short buffer and only allowed to shift the effective history if the next window confirms it. This rule is how the model separates a gift purchase from a new hobby. The $\lambda_k$ and the buffer threshold are trained on validation data with the backbone frozen, which isolates the module's contribution.

The interpretable output is a small set of drift curves with class sizes per dataset. We will check whether the same phenotype shapes recur across MovieLens-1M, Amazon Beauty, Amazon Toys, and Yelp.

Baselines

The naive reference is each backbone (SASRec and FlowRec) with no recency weighting, trained as published. This shows how much any forgetting helps at all.

The global-decay baseline adds one forgetting rate for everyone, tuned on validation. It is the simplest fix for preference inertia, and it is the bar the drift classes must clear for there to be any case for heterogeneity.

The strongest existing method is FlowRec from Preference Trajectory Modeling via Flow Matching for Sequential Recommendation. It is the most recent trajectory-style model in the file and was evaluated on exactly our four benchmarks. The paper reports a 9.87% relative improvement in HR@5 over its strongest baseline on Amazon Beauty and a 21.28% improvement over DiffuRec on long-tail items on Amazon Toys. No reported result was found on Lacuna for FlowRec's absolute HR@5 or NDCG values on any dataset, so we will reproduce it ourselves and report our reproduced numbers next to the published relative gains. Beating it means our FlowRec-plus-drift-classes variant improves NDCG@10 over reproduced FlowRec by more than the seed-to-seed standard deviation over five seeds, with the gain concentrated in the short-history and post-shift strata we predict.

Three ablations isolate the key ingredient, which is mixture pooling over drift trajectories. The per-user decay ablation estimates each user's forgetting rate from their own drift curve without pooling; it tests whether groups matter at all. The unimodal shrinkage ablation shrinks per-user rates toward a single population mean, as a mixed model would. This is the most dangerous competitor, and our claim stands or falls on beating it. The taste-cluster ablation forms $K$ groups by clustering users' mean item embeddings rather than their drift curves, then learns one decay per cluster. It tests whether what we group on matters.

For fairness, every variant uses the same frozen backbone checkpoint, the same temporal split, the same validation set, and an equal hyperparameter search budget of the same number of trials per method.

Why not just…?

Why not just tune a single global time decay? This may capture most of the overall accuracy gain, and we expect it to be competitive on the headline metric. The evidence that heterogeneity matters is indirect: the HbA1c cohort in Effect of Hemoglobin A1c Trajectories on Future Outcomes in a 10-Year Cohort With Type 2 Diabetes Mellitus shows that averages hide opposing subgroups. For recommendations, the global-decay baseline settles the question directly. If it matches our method in every stratum, drift heterogeneity is not practically important on these data.

Why not just estimate each user's forgetting rate individually, shrunk toward the population? This is the real threat. Hierarchical shrinkage already borrows strength for short-history users, and if drift rates are unimodal it is the correct model. Our method helps only when the distribution has separated modes, because then a short history that looks like a "settler" should be pulled toward the settler curve, not toward the grand mean. We therefore report the fitted mixture's BIC against the one-class model first. If $K=1$ wins, we expect, and will report, no gain over shrinkage.

Why not just add a mixture-of-experts layer to SASRec and let it learn groups? A gating network over several expert heads can in principle discover drift-based groups implicitly. Nothing in the file shows it cannot, and it may match our accuracy. We will run it as an extra comparison at a matched parameter count. We expect its gates, if clustered, to track taste rather than drift, and we can test this by measuring the mutual information of its gate assignments with our drift classes versus with the taste clusters. If the gates recover drift phenotypes and match accuracy, the contribution narrows to interpretability and the explicit noise-versus-shift rule, and we will say so.

Why not just run classical latent class growth analysis on the raw interaction logs? Rigid classes with zero within-class variance ignore the individual variation that Trajectories of Depressive Symptoms Among Web-Based Health Risk Assessment Participants found necessary. Fitting curves directly over high-dimensional item spaces is also where these estimators break down. Our compression into a low-dimensional drift signal plus random intercepts is exactly the response to this objection. We also include a zero-variance version of our model to check that the random intercept earns its place.

Proposed Experiments

All four benchmarks (MovieLens-1M, Amazon Beauty, Amazon Toys, Yelp) are evaluated with a global temporal split rather than leave-last-out, so the test period comes after training in calendar time and real drift is present. We report HR@10 and NDCG@10 overall and in three stratifications: by history length (bottom, middle, and top thirds); at sustained-shift events; and at transient events.

A sustained shift is a window whose drift exceeds the user's running 90th percentile and whose next two windows stay near the new region. A transient event exceeds the threshold but reverts. For sustained shifts we measure adaptation lag, the number of interactions until NDCG returns to the user's pre-shift level. For transient events we measure the NDCG drop on the following interaction.

There are three decisive checks. First, the drift distribution must be multimodal: BIC must favor $K>1$, and the classes must be stable, with adjusted Rand index at least moderate between time-split halves. Second, the drift-class model must beat unimodal shrinkage and taste clusters specifically on short-history users and on adaptation lag, while not worsening the transient-event drop. Third, the gain must hold on both backbones, which shows the module is not an artifact of one architecture.

A secondary check applies the model to AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics, where preferences shift within a dialogue. It asks whether session-scale drift shows the same class structure. No reported result was found on Lacuna for any method on AcCoRD, so this part is exploratory.