Recommender systems that track changing tastes have to decide how fast a user's preferences may move between two observations. Nearly all of them tie that speed either to wall-clock time, so that a month of silence counts as a month of drift, or to the count of interactions, so that the tenth click is one step after the ninth whatever the gap. Both choices are arbitrary, and both break on real users. People shop in bursts and then go quiet. Some silences are expected, such as nights, off-seasons, or the stretch before a replenishment purchase is due. Others are informative, as when a daily listener stops for three weeks. This note proposes that preferences drift on an operational clock: time measured by how much activity the user was expected to produce. Every expected-but-missing interaction then advances the clock, and quiet periods that were predicted to be quiet do not.
The mechanism comes from the time-rescaling theorem of point-process statistics. If one integrates an event stream's expected rate over time, the resulting "compensator" turns a bursty, irregular stream into a steady unit-rate one. We make that integrated rate the time axis on which a latent preference vector performs a random walk. Two difficulties follow. Most users have too few events to estimate their own rate, so the clock has to be learned jointly across users. The clock is also uncertain, so the model must carry that uncertainty into its beliefs about how far preferences have moved. We handle both with a hierarchical rate model and structured variational Bayesian inference, which approximates the intractable posterior by fitting a simpler distribution through optimization. The scientific question stays simple: which clock does preference drift actually run on, and does the answer change predictions after long absences?
Core Claim
The claim has two parts. Preference drift is better described as diffusion in the user's operational time than in physical time or step count. Estimating that operational time under extreme sparsity requires sharing statistical strength across users and propagating clock uncertainty into the drift posterior. The model nests all three clocks. For user $u$ with activity rate $\lambda_u(s)$, define
$$\tau_u(t) = w_{1}\,\frac{t}{T} + w_{2}\,\Lambda_u(t), \qquad \Lambda_u(t)=\int_0^t \lambda_u(s)\,ds,$$
where $T$ is a normalizing time scale and $w_1, w_2 \ge 0$ are learned weights with hierarchical priors. The user's preference vector $\theta_u$ then evolves as a Gaussian random walk in $\tau$:
$$\theta_u(\tau') \mid \theta_u(\tau) \sim \mathcal{N}\big(\theta_u(\tau),\, \sigma^2(\tau'-\tau) I\big).$$
Item choices at each event are drawn from a softmax over $\theta_u(\tau_u(t))^\top \beta_i$, where $\beta_i$ is the embedding of item $i$. The posterior over $w_2/(w_1+w_2)$ is directly interpretable. Near one means drift is driven by engagement. Near zero means it is driven by calendar time. Step-count time is the special case where $\Lambda_u$ jumps by one at each event and stays flat between events, and we fit it as an explicit competitor.
The idea combines pieces from several lines of prior work. The rescaling operation itself comes from work that maps event streams through a learned compensator for goodness-of-fit and anomaly testing (Detecting Anomalous Event Sequences with Temporal Point Processes; Conformal Anomaly Detection in Event Sequences). Those papers use rescaled time to test sequences. We use it as the time axis for a latent state. Analyzing Temporal Dynamics of Consumer's Behavior Based on Hierarchical Time-Rescaling is the closest use in consumer data. It builds a three-layer rate model (promotion responses, self-excitation, replenishment hazards) and rescales purchase timing per customer. However, it fits each customer and category independently by maximum likelihood and models when people buy, not what they come to prefer. Our proposal adds the collaborative pooling and posterior uncertainty that it explicitly lacks. Recurrent Poisson Factorization for Temporal Recommendation is the closest Bayesian recommender. It models user–item events as rate processes with exponential decay kernels in physical time and fits them with mean-field variational inference. Its baseline preferences $\theta_u^\top\beta_i$ are static, and its temporal effects are excitation kernels, not drift of the preference vector. Deterministic continuous-time recommenders motivate the uncertainty question from the other side. Long Short-Term Preference Modeling for Continuous-Time Sequential Recommendation finds that recency dominates short-term preference and that its memory is sensitive to noisy interactions. Learning Graph ODE for Continuous-Time Sequential Recommendation integrates drift through idle intervals with an ordinary differential equation solver. Both learn point estimates in physical time.
Research Program
We build a marked point-process model with two coupled parts.
The activity part gives each user a rate $\lambda_u(t)=\exp(b_u + g(t) + h_u(t))$. Here $b_u$ is a user baseline drawn from a population prior, $g$ is a shared weekly and seasonal profile, and $h_u$ is a self-exciting term that raises the rate for a while after each event. It uses the same ingredients as HTRm, but the parameters are pooled across users. A user with five events thus inherits a sensible clock from the population and is individualized only as far as the data allow. Because $g$ is piecewise constant and the excitation is exponential, $\Lambda_u(t)$ has a closed form, and no numerical solver is needed. That matters given the solver cost reported for GDERec.
The preference part is the random walk above, observed through item choices.
For inference we use a structured variational family rather than a fully factorized one. Each user's trajectory gets a Gaussian Markov posterior, fitted with a Kalman-style forward–backward pass inside a reparameterized evidence lower bound (ELBO, the standard variational training objective). This keeps correlations between consecutive preference states, which mean-field approximations such as RPF's discard and which the file identifies as a cause of overly narrow credible intervals. The clock parameters $(w_1, w_2, b_u, h_u)$ are sampled rather than plugged in. Each ELBO sample therefore draws a clock and then a trajectory given that clock. This is how clock uncertainty reaches the predicted preference after a long gap. Cost scales linearly in the number of events per user, as with RPF.
We know of one design risk from prior work. ContrastVAE: Contrastive Variational AutoEncoder for Sequential Recommendation shows that sequential latent-variable recommenders can ignore their latent variables, with the posterior collapsing to the prior. Our latent state has no expressive decoder that could bypass it, so we expect less risk. We will still monitor the per-user KL term. If collapse appears on the sparsest data, we will add a ContrastELBO-style agreement term between two perturbed views, treating that as a fallback rather than a contribution.
Baselines
The naive reference point is a static Bayesian factorization with no drift at all. We use Hierarchical Poisson Factorization as run in the RPF paper, which reports NDCG@20 of 0.18 on its Last.fm subset, against 0.51–0.53 for the RPF variants (Recurrent Poisson Factorization for Temporal Recommendation). This model tells us how much of any gain is due to temporal modeling at all.
The strongest competitor in the same probabilistic family is RPF itself. It asks the same question with Bayesian machinery, but it uses physical-time decay kernels and static preference vectors. We will rerun it on the exact Last.fm variant it used (1,200 users, 1,000 artists, about 480,000 events) and on the Tianchi click data. The RPF paper reports NDCG@20 of 0.51–0.53 and a returning-time mean absolute error of 165 for its DSRPF variant, against 280 for TSRS. We reproduce these numbers before comparing. Beating RPF means higher NDCG@20 and lower returning-time error on that split, plus better held-out event log-likelihood.
The strongest accuracy-oriented competitor is LSTSR. It reports a 71.87% NDCG@10 improvement over TGSRec and JODIE on Amazon Software, and a tenfold inference speedup (Long Short-Term Preference Modeling for Continuous-Time Sequential Recommendation). We also include GDERec, which reports average gains of 6.5% in Recall@5 and 8.8% in Recall@10 over SASRec and LightGCN (Learning Graph ODE for Continuous-Time Sequential Recommendation). These papers report only relative improvements, so no reported result was found on Lacuna for their absolute NDCG@10 or Recall@10. We will run both from their released code on identical chronological splits, with the same hyperparameter-search budget in GPU-hours as our model. Beating them on ranking alone is not the bar. The bar is matching their ranking accuracy within noise while clearly winning on calibrated return-time and post-gap prediction.
We add ContrastVAE as the variational discrete-sequence baseline. It reports Recall@40 gains of 8.86% to 19.34% over SASRec and S3Rec on Amazon Beauty, Toys, Tools and Office, but no reported absolute value was found on Lacuna. We also add VNTPP, the variational point-process baseline, which reports F1 of 0.596 against 0.503 for the Transformer Hawkes Process on Retweet (Variational Neural Temporal Point Process).
The key ablations isolate the clock while holding everything else fixed. We compare physical-time-only drift ($w_2=0$), step-count drift, a learned monotone function of the raw gap, and operational-only drift ($w_1=0$). A second ablation sets the clock to its point estimate instead of sampling it, and a third replaces the Markov posterior with mean-field. If the operational clock gives no gain over step count and a learned gap warp, the core claim fails.
Why not just…?
Why not just count interactions, as sequential models already do? Step count is a crude operational time, so this is a real threat. The two differ in exactly one situation: between events. Step count freezes the clock during a gap. Operational time advances it by the activity the user was expected to produce, so an unexpected three-week silence from a daily user counts as a lot of elapsed time, while a predicted weekend lull counts as little. The step-count ablation, evaluated on the first interaction after gaps binned by how surprising they were under the rate model, settles the question. If the two clocks tie even on the most surprising gaps, we will report that step count suffices.
Why not just learn a monotone warp of the gap length, such as log-time embeddings? A learned function of $\Delta t$ is the same for every user and every context. It cannot know that ten days is long for one user and normal for another, or normal in August and long in December. It also needs no point-process model, which makes it a fair and cheap competitor. We run it as an ablation with the same preference model. The operational clock earns its keep only if it wins on users whose activity rates are far from the population average.
Why not just add MC dropout or ensembles to LSTSR for uncertainty? Such intervals measure disagreement among fitted networks. They do not grow with elapsed operational time, so they cannot express "we are less sure because this user has been unexpectedly absent." LSTSR's reported sensitivity to noisy history suggests its point estimates move sharply after single accidental events. We will compare coverage of 80% predictive sets for the post-gap item and interval coverage for return time against a five-member LSTSR ensemble. If the ensemble calibrates equally well, the Bayesian treatment of the clock adds only interpretability, and we will say so.
Proposed Experiments
Synthetic data come first. We generate users with known rate processes and preference random walks run in each of the three clocks, at event densities that match the sparsest Amazon subsets. The decisive check is whether the posterior over $w_2/(w_1+w_2)$ recovers the true clock, and whether sampling the clock gives nominal interval coverage where the plug-in clock under-covers.
On real data we use the RPF Last.fm variant and Tianchi, ML-1M, Amazon Tools, Beauty and Software (sparse, long-tail), and Wikipedia and Reddit. All splits are chronological. Metrics are NDCG@k and Recall@k for the next item, return-time mean absolute error, held-out event log-likelihood, coverage of predictive sets and intervals, and a Kolmogorov–Smirnov test on rescaled inter-event times following Shchur et al. The KS test checks whether the pooled rate model gives a valid clock at all.
The central analysis is stratified. We sort users by how bursty their gaps are and by how many events they have, and we sort test events by how surprising the preceding gap was. The hypothesis predicts a specific pattern. Gains over step-count and physical-time drift should concentrate after surprising gaps and among bursty users. Gains should vanish on regular, dense users. The fitted clock weights should favor operational time in music listening and calendar time in seasonal e-commerce categories. A flat pattern across strata would refute the claim even if average accuracy improved.
References
- Detecting Anomalous Event Sequences with Temporal Point Processes
- Conformal Anomaly Detection in Event Sequences
- Analyzing Temporal Dynamics of Consumer's Behavior Based on Hierarchical Time-Rescaling
- Recurrent Poisson Factorization for Temporal Recommendation
- Long Short-Term Preference Modeling for Continuous-Time Sequential Recommendation
- Learning Graph ODE for Continuous-Time Sequential Recommendation
- ContrastVAE: Contrastive Variational AutoEncoder for Sequential Recommendation
- Variational Neural Temporal Point Process