A new user of a recommender and a long-standing user whose tastes have just changed are in the same position. In both cases the system must place that user in its learned preference space from a handful of fresh observations. The literature treats these as separate problems. Few-shot meta-learning handles new users on flat user-item matrices. Streaming tensor methods follow shifting structure but need many observations per entity. We propose one estimator for both, built on a multi-criteria tensor model: each rating is recorded along several axes (for a hotel, say value, location, service and cleanliness), and a shared "core" learned from the whole population describes how those axes combine. With the core and the item factors frozen, positioning a user is a small regression problem. What remains to learn is the starting guess for that regression: a population-level prior for newcomers, and a time-aware carry-over prior for returning users, plus a test that decides when a returning user has drifted far enough to be treated as new again.
The main claim is quantitative. When interactions are multi-criteria, each rated item contributes several roughly independent equations about the user rather than one. The number of interactions needed to position a cold or freshly drifted user should therefore fall by roughly the effective rank of the per-item criterion signals. This advantage is lost the moment the tensor is flattened into a matrix or a sequence. The claim can fail cleanly: if criteria ratings are nearly collinear, the gain disappears, and we will measure exactly that.
Core Claim
We use a Tucker decomposition, which factors a multi-way array into one small factor matrix per axis plus a small core array that mixes them. With user factor $\mathbf{u} \in \mathbb{R}^P$, the predicted rating of item $i$ on criterion $c$ is
$$\hat{x}_{uic} = \mathbf{u}^\top \mathbf{w}_{ic}, \qquad \mathbf{w}_{ic} = \mathcal{G} \times_2 \mathbf{v}_i \times_3 \mathbf{z}_c ,$$
where $\mathcal{G}$ is the core, $\mathbf{v}_i$ and $\mathbf{z}_c$ are the item and criterion factors, and $\times_n$ multiplies the core along its $n$-th axis. With these fixed, $\mathbf{w}_{ic}$ is a known regressor. Under a Gaussian prior $\mathbf{u} \sim \mathcal{N}(\mathbf{m}, \mathbf{S})$ and noise variance $\sigma^2$, the estimate from observed cells $\Omega$ is closed form:
$$\hat{\mathbf{u}} = \Big(\mathbf{S}^{-1} + \sigma^{-2}\sum_{(i,c)\in\Omega} \mathbf{w}_{ic}\mathbf{w}_{ic}^\top\Big)^{-1}\Big(\mathbf{S}^{-1}\mathbf{m} + \sigma^{-2}\sum_{(i,c)\in\Omega}\mathbf{w}_{ic}\,x_{uic}\Big).$$
The data term gains rank as quickly as the regressors $\{\mathbf{w}_{ic}\}_c$ for a single item span new directions. That is the source of the predicted gain: error curves plotted against $K$ rated items should shift left by about the effective rank $r_{\text{eff}}$ of those per-item regressors, capped at $C$ criteria.
The contribution lies in the prior. For new users, $(\mathbf{m}, \mathbf{S})$ is chosen from a learned mixture of population components, selected by predictive likelihood on the first observations. For a returning user last seen with estimate $\mathbf{u}_{\text{prev}}$ a time $\Delta t$ ago, the prior is a carry-over $\mathcal{N}(\mathbf{A}\mathbf{u}_{\text{prev}}, \mathbf{S}_0 + \Delta t\,\mathbf{Q})$, where $\mathbf{A}$ and $\mathbf{Q}$ are learned from population trajectories. A Bayes-factor test on the most recent observations compares the carry-over prior against the population mixture. When the mixture wins, the user is reset. We call this a "re-cold-start", the regime the literature flags as unresolved for sequential models.
Relation to Prior Work
The tensor side supplies the model class, but none of it positions new or drifting users. Deep Transfer Tensor Decomposition with Orthogonal Constraint for Recommender Systems (DTTD) uses Tucker with autoencoder side features to handle more than 99% sparsity on TripAdvisor and RateBeer, and aligns cores across domains. It is static and offline, and we reuse its data regime and core structure. Coupled Graphs and Tensor Factorization for Recommender Systems and Community Detection positions cold users through a side graph, with user factors that do not change over time. We need no side graph, and we let factors move. Bayesian Poisson Tucker Decomposition for Learning the Structure of International Relations gives proper probabilistic treatment of empty cells and a Tucker core over time steps, but only for population-level event modelling. Its Bayesian view supports our closed-form posterior.
Two streaming tensor methods, Identifying and Alleviating Concept Drift in Streaming Tensor Decomposition and Multi-Aspect Streaming Tensor Completion, motivate our drift test. SeekAndDestroy tracks the birth and death of latent concepts by matching factor correlations against a library. OR-MSTC completes growing tensors robustly. Both require dense batches per entity, and neither handles an individual user with near-zero post-drift data. Our reset test moves SeekAndDestroy's population-level "new concept versus persisting concept" decision down to the individual user.
The meta-learning side constrains the design. MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation and Preference-Adaptive Meta-Learning for Cold-Start Recommendation show that one global starting point is too rigid. MAMO answers with profile memories and PAML with graph-path gating. Our mixture prior addresses the same rigidity without demographics, which MAMO needs, and without graph path search, which PAML needs. Sequential Recommendation for Cold-start Users with Meta Transitional Learning shows that lightweight models beat Transformers when $K \le 4$. That result supports a closed-form regression over a deep adapter. It also sets a bar: flattening context into a sequence must lose to our model before the tensor structure earns its place.
Research Program
We will first fit the Tucker population model (core, item factors, criterion factors) on warm users under forward-in-time splits. We will then learn the prior components from the posterior distribution of warm users' factors, and learn $(\mathbf{A}, \mathbf{Q})$ from consecutive-window factor estimates of long-history users. Next, we will measure $r_{\text{eff}}$ per dataset from the singular spectrum of the stacked per-item regressors. This gives a prediction of the gain made before any test data is seen. Finally, we will evaluate cold start and drift recovery under a single protocol that reveals $K$ items one at a time. The whole adaptation step is a $P \times P$ solve per user, so compute is not a confound.
Baselines
The naive reference points are item-mean prediction per criterion, and a "do nothing" carry-over that keeps a returning user's old factor forever. Beating these shows only that personalization happens at all.
The key ablation is plain fold-in: the same frozen Tucker model and the same closed-form solve, but with a single isotropic prior. It is the strongest cheap competitor, and it isolates whether the learned priors and reset test matter. A second ablation fits the identical pipeline to the overall-rating matrix alone. This isolates the multi-criteria contribution that the core claim predicts.
For warm, static accuracy the strongest tensor method is DTTD, which reports RMSE 0.876 on TripAdvisor-12M against 0.922 for the next-best baseline DTF (DTTD). That result comes from a static split, so we will rerun DTTD on our time-ordered cold-start splits rather than compare against the published figure. For cold start, the strongest meta-learners are MAMO, which reports MAE 0.8894 against MeLU's 0.9299 in the cold-user, cold-item setting of MovieLens-1M (MAMO), and PAML, which reports only relative RMSE gains of 1.9–5.3% on Douban Book and 4.1–6.9% on Yelp. Both will be applied to the tensor by giving them criterion-conditioned output heads. For drift, we run MetaTL on context-augmented sequences and OnlineCP-style streaming updates with SeekAndDestroy rank tracking.
All methods get the same splits, the same latent dimension budget, and an equal hyperparameter-search budget on a time-held-out validation window. Our method counts as better only if it reduces error in the first one to five post-arrival or post-drift items relative to both plain fold-in and the best meta-learner, and does not lose warm RMSE relative to DTTD.
Why not just…?
Why not just use plain fold-in with a ridge penalty? This is the real threat. It may capture most of the gain, since the multi-criteria effect lives in the data term, not the prior. The fold-in ablation settles it directly. If learned priors add nothing at $K \ge 2$, we will report the contribution as the multi-criteria identifiability result together with the reset test. The reset test cannot be replaced by any fixed ridge penalty.
Why not just meta-learn the tensor model with MAML-style fine-tuning? Our posterior is exactly what a linear head reaches after unlimited inner-loop steps under a Gaussian prior. MAML approximates it with a few noisy gradient steps. The MAMO and PAML reruns with tensor heads will show whether their extra machinery beats the exact solve. MetaTL's finding that simpler adapters win at small $K$ suggests it will not.
Why not just feed criteria and timestamps as features to a sequential recommender? A sequence model can in principle learn criterion interactions. But it spends its scarce few-shot data learning them per user, where our model inherits them from a population core. The context-augmented MetaTL baseline tests this directly, and the matrix-only ablation shows how much structure flattening throws away.
Why not just down-weight old data with exponential time decay? Decay trades stability for plasticity with one global rate, which is the dilemma the literature describes. Our carry-over prior widens uncertainty with elapsed time and resets only when the evidence calls for it. We will tune a decayed fold-in baseline to its best rate and compare recovery curves.
Proposed Experiments
We use TripAdvisor-12M (4 criteria, 99.12% sparse), TripAdvisor-20M (8 criteria, 99.89% sparse) and RateBeer (5 criteria), ordered by review date. That review timestamps are retained in the processed versions must be verified first. MovieLens-1M serves as a matrix-only sanity check against MAMO's setting. Metrics are per-criterion and overall RMSE and MAE, NDCG@10, and area under the error-versus-$K$ curve for $K$ from 1 to 10.
The first decisive check tests the core prediction. Across the three datasets, the leftward shift of the all-criteria curve over the overall-only curve should track the measured $r_{\text{eff}}$. That means larger for 8-criterion TripAdvisor than for 4-criterion, and near zero wherever criteria turn out to be collinear. If no such relationship appears, the core claim is wrong.
The second check tests drift recovery. On natural data, we measure post-reset error for users the test flags. On a semi-synthetic set with known change points, we splice two real users' histories together. There, reset detection is scored by delay and false-alarm rate, and recovery by error over the first five post-change items, against carry-over, decayed fold-in and streaming CP.
The third check asks whether this whole class of simple regressions is too weak. It compares warm-user RMSE against the rerun of DTTD, to confirm that freezing the core for fast adaptation does not cost accuracy on established users.
References
- Deep Transfer Tensor Decomposition with Orthogonal Constraint for Recommender Systems
- Coupled Graphs and Tensor Factorization for Recommender Systems and Community Detection
- Bayesian Poisson Tucker Decomposition for Learning the Structure of International Relations
- Identifying and Alleviating Concept Drift in Streaming Tensor Decomposition
- Multi-Aspect Streaming Tensor Completion
- MAMO: Memory-Augmented Meta-Optimization for Cold-start Recommendation
- Preference-Adaptive Meta-Learning for Cold-Start Recommendation
- Sequential Recommendation for Cold-start Users with Meta Transitional Learning