One Filter for New and Returning Users: Cold Start as a Preference Jump

A recommender meets the same problem in two forms. A brand-new user arrives with no history. A long-time user returns after months away, or suddenly starts buying in an unfamiliar category. Current systems treat these as separate problems. Meta-learning methods handle the new user, and temporal models handle the drifting one. For the returning user they force a crude choice: reset the profile and discard years of evidence, or trust the old profile and keep recommending what the person no longer wants. We propose treating both cases as one inference problem. Each user's taste is a hidden state that moves continuously over time. With some probability, the state can "jump" to a fresh draw from the population of users active at that moment. A new user is simply a user whose jump probability is one at entry. A returning user has a jump probability that grows with the length of the absence.

The concrete claim is that a continuous-time tensor factorization with this jump component can do two things. First, it can match specialized meta-learners on true cold starts. Second, it can clearly beat both meta-learners and smooth temporal models on returning and domain-shifting users. The gain comes from two sources. The model conditions on when the user arrived: a 2024 newcomer resembles other 2024 newcomers more than the all-time average user. And the model reports, per user, a posterior probability that their taste has reset. That probability is an interpretable, testable diagnostic for the "cold start or drift?" ambiguity that the literature names but does not resolve.

Why this intersection, and what is already done

Temporal tensor factorization stores interactions as a multi-way array (users × items × time) and learns a low-dimensional vector per entity, plus a description of how those vectors change over time. The most relevant models make that change continuous. Bayesian Continuous-Time Tucker Decomposition lets the tensor's interaction core vary through a Gaussian process. It rewrites the process as a linear stochastic differential equation so inference costs linear time through Kalman filtering. SONATA streams factor trajectories through Matérn-kernel state-space models and flags "novel" entities for its data-selection step. Functional Complexity-adaptive Temporal Tensor Decomposition uses neural ODEs for the trajectories.

These papers supply our inference machinery, and they also motivate the gap. All three are transductive: an entity with no observed slices has no factor. All three assume stationary, smooth dynamics, which over-smooth abrupt reversals of taste. SONATA's novelty score decides which cold entities to ingest. It does not decide what to believe about them before data arrives. SONATA and BCTT also assume Gaussian observations, which do not fit clicks and views.

The closest prior work is Dynamic Tensor Recommender Systems. It fits spline-based time-varying tensor factors and handles new items by borrowing strength from static product subgroups. Our proposal differs in three ways. The borrowed prior is a function of entry time, not a fixed hierarchy. The user side is covered, not only the item side. And sudden change is modeled explicitly rather than smoothed by splines.

On the cold-start side, MeLU and MAMO learn initializations that adapt to a new user within a handful of ratings. MAMO adds memories because a single global starting point fits diverse users poorly. PAML personalizes that starting point through graph neighbors. MetaTL and Mecos extend meta-learning to short sequences. All of them treat the user's taste as fixed during adaptation and ignore elapsed wall-clock time.

On the drift side, two papers separate long-term from short-term taste in different ways. Towards Explainable Conversational Recommendation keeps a static "global view" of long-term taste and a fast "local view" for the current session. PESO keeps one evolving adapter with a proximal penalty that lets well-evidenced directions move and anchors the rest. Neither can initialize a user without history. Neither says whether a sparse burst is exploration or a real change.

Core Claim

The obvious combination is "meta-learn the initial state of a continuous-time tensor model." On inspection we think it is mostly redundant. In a state-space model, a new user with a good prior already adapts in closed form: a Kalman update on five observations is exact few-shot Bayesian adaptation. So an outer meta-learning loop over gradient steps adds little beyond what a well-chosen prior gives. The open question is therefore not how to adapt quickly but which prior to adapt from, and when to fall back to it. That is what we build.

Let $u(t)$ be a user's taste vector at time $t$. Let $m(t)$ and $\Sigma(t)$ be the mean and covariance of taste vectors among users entering or active at time $t$. Let $\Delta$ be the time since the user's last observed interaction. The transition is a two-component mixture:

$$u(t+\Delta) \sim (1-\pi_u(\Delta))\,\mathcal{N}\big(a(\Delta)\,u(t) + (1-a(\Delta))\,m(t+\Delta),\ \sigma^2(\Delta) I\big) \;+\; \pi_u(\Delta)\,\mathcal{N}\big(m(t+\Delta),\ \Sigma(t+\Delta)\big).$$

Here $a(\Delta)=e^{-\lambda\Delta}$ is the persistence of an Ornstein–Uhlenbeck-style diffusion toward the population mean. $\sigma^2(\Delta)$ is its accumulated variance. $\pi_u(\Delta)$ is a jump hazard predicted from the gap length and user features, with $\pi_u=1$ at first entry. Item vectors evolve by slow diffusion only. A rating or click is modeled with a logistic or ordinal likelihood on the inner product of user and item vectors at the interaction time, updated by expectation propagation as in BCTT.

The population trajectory $(m(t), \Sigma(t))$ is itself a smooth state-space process. It is fit to the inferred entry states of existing users. This is how "who joins when" enters the model.

The object the method returns is the filtered posterior over $u(t)$ together with the posterior jump probability. That probability is the explicit answer to "did this person change, or do we just lack data?"

Research Program

The build has three parts.

The first is a CP-style temporal tensor model with the mixture transition above. We avoid a Tucker core to sidestep the $R^K$ growth noted for BCTT. Filtering uses a two-component Gaussian-sum filter collapsed by moment matching after each observation, which keeps cost linear in the number of events.

The second is learning the hazard function and the cohort prior $(m(t), \Sigma(t))$ by maximizing the filter's one-step-ahead predictive likelihood on chronologically earlier users.

The third is an analysis track that asks whether the posterior jump probability is meaningful. The test is whether it predicts, on held-out future interactions, which of two forecasts wins: the user's pre-gap state, or the population prior at return time.

We also add a discrete observation model for implicit feedback. We will use the randomized-exposure slices of KuaiRand to check that inferred jumps are not artifacts of exposure policy. Otherwise a "change" could simply reflect what the platform stopped showing, the confound documented by KuaiRec.

Baselines

The naive reference points are two. The first is a recency-weighted popularity recommender computed over the current time window. The second is a matrix factorization retrained on a sliding window, with new users given the mean factor. Both use the same chronological splits. Beating them by a clear margin on cold and returning users is the minimum bar. KuaiRec's finding that popularity looks strong on biased logs but collapses on fully observed data means we also report these baselines on KuaiRec's dense matrix.

The strongest cold-start competitor is MAMO. MAMO reports MAE 0.8894 on MovieLens 1M in the cold-user, cold-item setting, against 0.9299 for MeLU. We will first reproduce MAMO on its own protocol to confirm our implementation. We then rerun it on our chronological protocol, with timestamps and gap lengths added as input features so it is not handicapped on time. Tuning budgets are matched in GPU-hours and in hyperparameter trials. On true new users, we count parity with MAMO (within its seed variance) as success. On returning and domain-shifting users, success means a significant MAE or NDCG improvement.

For sequential cold start we rerun MetaTL. The file reports only an 11.7% relative gain over SASRec and BERT4Rec, not an absolute value. We rerun it with the same support sizes of 2 to 4 on the Amazon categories it used.

For drift we run BCTT. No reported result was found on Lacuna for BCTT beyond qualitative claims of lower RMSE on MovieLens 100K. We give it mean-factor initialization for unseen users. We also run PESO's proximal-regularization idea transplanted onto our factor model, a KL-anchored update of user vectors, to isolate whether gated stability alone explains any gain.

The key ablations remove one ingredient at a time. The first removes the jump component ($\pi_u=0$ after entry). The second replaces the entry-time prior with a time-invariant global prior. The third does both, which yields a plain continuous-time Kalman tensor model.

Why not just…?

Why not just meta-learn the initialization with MAML-style methods and add timestamps as features? This is the most natural competitor, and on true new users it may tie us. That is acceptable, because our claim there is parity, not dominance. Where it should fail is the returning user. A meta-learned initializer must either restart from the global prior, discarding history, or fine-tune from the stale profile. It has no principled way to weigh the two by gap length. The returning-user split will show whether that matters. If MAMO with gap features matches us there too, the jump mechanism is not earning its keep, and we will say so.

Why not just run a continuous-time Kalman tensor model with a global prior, as BCTT or SONATA would? This is the real threat, and it is exactly our third ablation. Diffusion already inflates uncertainty with elapsed time, so a returning user's state naturally relaxes toward the mean. The difference is shape. Diffusion shrinks the old state gradually. A jump keeps the old state intact with probability $1-\pi$ and discards it otherwise, which yields a bimodal posterior. The recurrent tastes and abrupt resets described in AcCoRD should favor the bimodal form. If the ablation matches the full model, we will simplify to it and keep only the cohort prior.

Why not just borrow strength from static subgroups, as DTRS does? Subgroup borrowing fixes the prior by category or segment. We expect cohort effects over time on platforms whose user base changes, and a fixed hierarchy cannot capture them. The comparison between the entry-time prior and a feature-conditioned static prior tests this directly. Datasets with a stable user base may show no difference, which would bound the claim to growing platforms.

Why not just use a sliding window with popularity for new users? Windowing loses entity identity across windows, and DTRS already reports RMSE 12.504 against 16.571 for BPTF on IRI data, a gap attributed to discarding multi-mode temporal structure. Popularity is known to be inflated on biased logs. We include both anyway as floors.

Proposed Experiments

We use MovieLens 1M, following the MeLU and MAMO protocols but with chronological splits on rating timestamps, plus the Amazon Electronics and Movies & TV sequences used by MetaTL and PESO. Users are partitioned into four groups: true new users, returning users with a gap above a threshold such as 90 days, users entering a new category, and steady users. We report MAE and RMSE for ratings, and HR@10 and NDCG@10 for ranking, per group, with seeds and confidence intervals.

The first decisive check is the returning-user group. The full model must beat the three ablations, MAMO-with-time, and BCTT there. It must also stay within noise of MAMO on new users.

The second decisive check is calibration of the jump posterior. We bin returning users by inferred jump probability. In each bin, we measure how often the population-prior forecast beats the pre-gap forecast on the user's next ten interactions. A monotone relationship means the model genuinely separates drift from missing data. A flat one means it does not, whatever the headline metrics say.

The third check tests the cohort prior. On MovieLens and Amazon, we measure whether the entry-time prior lowers first-five-interaction error relative to a time-invariant prior. The effect should grow with the time distance between training cohorts and the test cohort.

The fourth check repeats the jump analysis on KuaiRand's randomly exposed items. If inferred jumps vanish when exposure is randomized, they were platform artifacts rather than preference change. That would be a result worth reporting in its own right.