Most recommenders read a user's history as an ordered list of items and ask what comes next. The timing is thrown away, or used only to down-weight old items. Yet the time a person returns is itself a strong clue about why they returned. Someone who comes back two days after buying a laptop is probably still in an "accessories" mood. Someone who returns after exactly a month may be restocking a consumable. Someone who returns after half a year has likely left both moods behind. We propose a sequential recommender that keeps several separate intent states per user. Each intent carries its own clock: its own rate of forgetting, its own growing uncertainty, and its own tendency to bring the user back to the platform. When a visit happens, the model asks which intent most plausibly caused a return at this moment and for this item. It updates mainly that intent, and at recommendation time it weights intents by how well each explains the current elapsed gap.
The opportunity sits between two literatures that have not been joined. Recommendation work has models that give each user several intent vectors and treat clicks as uncertain, but they step through interactions by index and ignore physical time. Irregular time-series work has models whose hidden state evolves in continuous time, but they assume observation times carry no information about the hidden state. In recommendation logs that assumption is plainly false, because users arrive because some desire became active. Our central claim is that modelling arrival time as evidence of which intent caused the visit improves next-item ranking, with the largest gains for users whose return gaps are long or vary widely. The claim is falsifiable, and a simpler auxiliary-loss variant could erase it. We test that directly.
What exists and what is missing
Five papers already join continuous time and recommendation, and none do what we propose. DeepCoevolve models both what item a user picks and when, using a point process (a model of event arrival rates over time) driven by recurrent user and item embeddings. This is the closest precedent for treating timing as a modelling target. Its states stay constant between events, however, and each user has one vector. GDERec and TGODE let embeddings drift through gaps using graph neural ODEs (networks whose hidden state follows a learned differential equation). GDERec uses a first-order Taylor closed form; TGODE uses a Runge-Kutta solver plus a diffusion model that invents plausible interactions during dormant periods. SS4Rec ties the step size of a state space model (a linear recurrent system that can be run in continuous time) to the physical gap and stacks a selective Mamba-style layer on top. IntervalLLM feeds MLP-embedded intervals into LLaMA-2 attention for "interval cold-start" users, meaning users who return rarely after long gaps. All five treat the gap as an input that shapes a single user state. None treats the gap as an observation that must be explained, and none attributes a visit to one of several intents.
The multi-intent side is represented by Unified Representation Learning for Multi-Intent Diversity and Behavioral Uncertainty. It keeps several attention-combined intent vectors and Gaussian uncertainty over interactions through a sequential variational autoencoder, but it operates over index steps. TIARec separates temporary interests from noise using multi-agent reinforcement learning, also without physical time. These papers motivate our design: they show intents and noise matter, and they leave time out.
Two papers constrain our design. Neural Jump ODEs prove convergence to the optimal predictor only when observation times are independent of the state. That is exactly the assumption we drop, so their guarantee does not transfer and we do not claim one. TGODE's gap-filling constrains us in a different way. Inventing interactions for a user who is genuinely dormant may inject false evidence. We take the opposite stance: during a gap, uncertainty grows and nothing is imputed. The comparison between the two stances becomes an experiment rather than an assumption.
Core mechanism
Each user carries $K$ intents. Intent $k$ has a Gaussian state with mean $\mu_k(t)$ and variance $\Sigma_k(t)$. Between visits, each intent follows a mean-reverting linear process. Its mean relaxes toward a slowly updated long-term anchor $a_k$ at a learned diagonal rate $\alpha_k$, and its variance grows toward a stationary level. Both are available in closed form, so no ODE solver is needed. This is the same computational trick that makes diagonal state space models cheap.
Each intent also has a return hazard, the instantaneous rate at which that intent brings the user back:
$$\lambda_k(t) = \mathrm{softplus}\big(w_k^\top \mu_k(t) + b_k + c_k^\top \phi(t)\big).$$
Here $w_k, b_k$ are learned, and $\phi(t)$ is a small set of sine and cosine features for time of day and day of week. These features let intents have daily or weekly rhythms.
The model treats the intents as competing causes of the next visit. Suppose the previous visit was at $t_0$ and the user next clicks item $i$ at time $t$. The event likelihood is
$$p(i, t) = \Big[\sum_{k=1}^{K} \lambda_k(t)\, p(i \mid \mu_k(t), \Sigma_k(t))\Big] \exp\Big(-\int_{t_0}^{t} \sum_{k} \lambda_k(s)\, ds\Big).$$
In this expression, $p(i \mid \mu_k, \Sigma_k)$ is a softmax over item embeddings, with an expectation over the Gaussian taken by a single reparameterized sample. The integral is computed with a few quadrature points, because the hazards are smooth between events.
The responsibility of intent $k$ for the visit is $r_k \propto \lambda_k(t)\, p(i \mid \mu_k(t), \Sigma_k(t))$. Each intent then receives a Kalman-style correction scaled by its responsibility: a large correction for the intent that explains the visit, almost none for the others. At recommendation time we score candidates at the query time. Intents are weighted by their current hazards, and the survival term penalizes intents whose rhythm predicts the user should already have returned.
Three consequences follow, and each is testable. First, the ranked list depends on when the user asks, possibly in a non-monotone way. Second, a long gap broadens uncertainty instead of forcing a confident drifted guess. Third, a click that is poorly explained by every intent's hazard gets diffuse responsibility and therefore a small update. This gives a principled treatment of accidental or exploratory clicks, which neither SS4Rec nor GDERec has.
Research program
We will build the model in PyTorch on top of an open SS4Rec-style codebase so that embedding tables, sequence batching and evaluation are shared. There are three stages.
First, we validate on synthetic users generated with two or three intents of known return rhythms and item preferences, some with informative gaps and some without. This checks whether responsibilities recover the true generating intent. It also checks whether the gain disappears, as it should, when gaps are made uninformative.
Second, we run the main benchmarks with a fixed tuning budget per method.
Third, we analyze where gains occur. We stratify users by the spread of their log-gap distribution and by IntervalLLM's warm/cold split, and we inspect the learned rates $\alpha_k$ and hazards for interpretable timescales.
We treat $K$ as a hyperparameter in $\{1, 2, 4, 8\}$. $K=1$ is itself a meaningful baseline: a single-intent continuous model trained with a time likelihood.
Baselines
The naive reference is a popularity-by-recency ranker together with SASRec, which ignores time entirely. Any continuous-time model must clear these, and the gap to SASRec calibrates how much time matters on each dataset.
The time-as-decay reference is TiSASRec, which adds learned interval terms to self-attention. IntervalLLM reports that TiSASRec degrades on interval cold-start users. We rerun it under our splits rather than relying on that claim.
The strongest continuous-time competitor is SS4Rec. It is the only prior model with linear cost in sequence length and physical-time step sizes, and its diagonal state already contains many timescales. That makes it the most direct test of whether our explicit intents add anything. The file contains no reported SS4Rec number, only statements that it beats SASRec, BERT4Rec and Mamba4Rec on MovieLens-1M, KuaiRec and Amazon Sports/Video Games. We will therefore run it ourselves with the authors' configuration and give it the same hyperparameter search budget, measured in trials and GPU-hours, as our model.
The strongest multi-intent competitor is the unified multi-intent/uncertainty model of Xu et al.. On Amazon Electronics it reports HR@10 of 66.4% and NDCG@10 of 47.2%, and HR@10 of 42.3% on users with a single action. We will reproduce it on Electronics under their protocol to confirm we match these figures before comparing on other datasets.
GDERec and TGODE are included on the Amazon and MovieLens subsets they used. The file gives only relative gains for them: 6.5% Recall@5 and 8.8% Recall@10 over SASRec/LightGCN for GDERec, and 10–46% over SASRec/DreamRec for TGODE. These show continuous models help, but they are not absolute targets.
IntervalLLM reports 61.7% HR@1 on Amazon Video Games under a 20-candidate multiple-choice protocol, against 56.0% for plain LLaMA-2. We will evaluate our model under that exact protocol on Video Games, CDs and Books, as a cost-accuracy comparison against a 7B model.
The key ablations isolate the proposal's ingredient. In the first, we keep $K$ intents and continuous decay but drop the hazard and survival terms, so time shapes the state but is never explained. In the second, we keep the time likelihood with $K=1$. In the third, we replace growing variance with deterministic decay.
Beating the field means three things. On NDCG@10, the model must significantly exceed SS4Rec and the Xu et al. model, using paired bootstrap over users on at least four of the six recommendation datasets. The gain must be concentrated, beyond chance, in the high gap-variance and cold-interval strata. And both ablations must lose most of that stratified gain.
Why not just…?
Why not just add a time-prediction loss to SS4Rec? This is the most serious threat. A single-state model trained to predict when the next visit happens, as DeepCoevolve does with a single vector per user, may absorb much of the benefit, because SS4Rec's diagonal state already mixes fast and slow timescales. Our $K=1$ time-likelihood ablation is essentially this model with a cleaner formulation, and we will also bolt a log-normal gap head onto SS4Rec directly. If either matches the full model on the high gap-variance stratum, the multi-intent attribution adds nothing. We would then report the result as "endogenous timing matters, intents do not" and drop the mixture.
Why not just feed bucketed or MLP-embedded gaps into a multi-intent model? Appending interval features to the Xu et al. architecture gives intents access to timing without any point-process machinery. The evidence we have cuts only partly against this. IntervalLLM shows that continuous interval embeddings beat raw timestamps in text prompts, but it never tested intent-specific hazards. We run this variant with a log-bucket embedding and with an MLP interval embedding. Our distinctive prediction is the one this variant cannot easily make: rankings that shift with the query time for a fixed history. We measure that shift directly, and the variant must also lose on the survival-based ranking when the query time is moved.
Why not just fill gaps with generated interactions, as TGODE does? TGODE reports large gains, so imputation clearly helps on its benchmarks. We have no evidence yet that it hurts. Our hypothesis is that it helps on dense, trend-driven data and harms users who are genuinely dormant. We compare TGODE against our variance-broadening model on the long-gap cold stratum, separately from warm users. If TGODE also wins there, our treatment of dormancy is wrong.
Why not just use an LLM with timestamps? Plain LLaMA-2 prompting reached 56.0% HR@1 on Video Games, against 61.7% for IntervalLLM's numeric interval embedding, which shows text timestamps are weak. Even the numeric approach needs a 7B backbone. We report accuracy together with serving latency on the same hardware.
Proposed experiments and decisive checks
The data are Amazon Video Games, Electronics, Beauty and Sports under five-core filtering, MovieLens-1M, KuaiRec Small and Large for long dense histories, and Tianchi TMALL for bursty, noisy sessions. IPTV, Yelp and Reddit are used for joint what/when evaluation against DeepCoevolve's MAR and event-time MAE. We apply a chronological leave-last-out split and report HR@10, NDCG@10 and Recall@10 on full-catalog ranking. On Video Games we additionally report HR@1 under IntervalLLM's protocol. We also report the MAE of the predicted return time and training and inference wall-clock time.
Three checks decide the proposal. The first is synthetic intent recovery: responsibilities should match true intents well above chance when gaps are informative, and the advantage over the no-hazard ablation should vanish when gaps are shuffled. The second is query-time sensitivity on real data. We re-score each test user's candidates at counterfactual query times. The model's ranking should change, and its accuracy at the true return time should exceed its accuracy at shifted times by more than SS4Rec's does. The third is the gain-concentration test: improvements over SS4Rec should grow with the spread of the user's log-gap distribution. If gains are flat across strata, whatever the model is doing, it is not the mechanism claimed here.
References
- Recurrent Coevolutionary Latent Feature Processes for Continuous-Time Recommendation
- Learning Graph ODE for Continuous-Time Sequential Recommendation
- Time Matters: Enhancing Sequential Recommendations with Time-Guided Graph Neural ODEs
- SS4Rec: Continuous-Time Sequential Recommendation with State Space Models
- Not Just What, But When: Integrating Irregular Intervals to LLM for Sequential Recommendation
- Unified Representation Learning for Multi-Intent Diversity and Behavioral Uncertainty in Recommender Systems
- Learning From Atypical Behavior: Temporary Interest Aware Recommendation Based on Reinforcement Learning
- Neural Jump Ordinary Differential Equations: Consistent Continuous-Time Prediction and Filtering