acceptodds
Under review as a conference paper at ICLR 2027

Matrix-Variate Gaussian Process State-Space Model with Structured Latent Dynamics

Abstract

Matrix-valued time series arise in many applications with coupled row-wise, column-wise, and temporal dependencies. Probabilistic modeling of their dynamics requires a framework that preserves row–column structure while remaining tractable for high-dimensional data. In this paper, we formulate a matrix-variate Gaussian process state-space model (MV-GPSSM) with structured latent dynamics. It represents the observed sequence as low-dimensional latent matrices through a Tucker2-type decomposition, models their temporal evolution using a matrix-variate GP, and lifts the predictive distribution back to the observation space through learned row and column bases. Within this formulation, we characterize how latent interaction mechanisms induce structured matrix-valued kernels: multiplicative and additive row–column interactions lead to Kronecker-product and Kronecker-sum covariance forms, respectively. Thus, these covariance structures arise as consequences of the assumed latent interaction mechanisms rather than being prescribed directly, thereby providing a dynamical interpretation of these structured covariance forms within the model family. We further present a computational complexity analysis of inference in the proposed MV-GPSSM. Experiments on battery electric vehicle fleet data and publicly available weather data show that the MV-GPSSM maintains predictive performance comparable to that of the baseline methods in terms of both point accuracy (RMSE) and probabilistic quality (NLL evaluated using the full predictive covariance matrix), while remaining tractable at large matrix sizes where observation-space methods become computationally or memory infeasible. Controlled comparisons clarify the roles of matrix structure, latent representation, and GP-based dynamics. The learned row- and column-wise covariance structures further reveal qualitatively interpretable dependencies in vehicle behavior and weather spatiotemporal patterns.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
matrix-variate Gaussian process, Gaussian process state-space model, matrix-valued time series, structured kernels, probabilistic prediction, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tKtNOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Matrix-valued time series arise in many settings where each time step is naturally represented as a matrix. Matrix-variate Gaussian processes (in short, MVGPs) provide a principled probabilistic framework for this task. However, direct MVGP prediction in the observation space is computationally expensive.

To overcome this bottleneck, this paper formulates a matrix-variate Gaussian process state-space model (MV-GPSSM) with structured latent dynamics.

The main idea is to represent the observed sequence as a third-order tensor and apply a Tucker2-based decomposition, which yields a sequence of low-dimensional latent matrices. Technically, this shifts the dominant cost from the observation space to the reduced latent space while preserving row–column structure and uncertainty-aware prediction. Within this formulation, the latent interaction structures induce Kronecker-product and Kronecker-sum covariance forms, respectively, providing a dynamical interpretation of these structured covariance forms in the proposed model family.

Experiments on BEV fleet data and publicly available weather data show that the proposed method maintains predictive performance comparable to that of the baseline methods in terms of both RMSE and NLL, while remaining tractable at large matrix sizes where observation-space methods become infeasible, including experiments with up to 800 vehicles.

Critical strengths

None.

Critical weaknesses

The two used datasets are quite far from public familiarity. And for these two datasets, what are the current performances of large models or deep neural networks? That should be adopted here to give comparisons, too.

Questions for authors

For these two used datasets, what are the current performances of large models or deep neural networks?

1 comment
AuthorsOpenReview ↗

Response to Reviewer tKtN

We thank the reviewer for the comments.

Q. How do large models or deep neural networks perform on the two datasets?

We did not run large models or deep neural networks on these datasets, because they do not provide the joint predictive distribution that we evaluate (General Response (5) Deep forecasters). Our target is the joint predictive distribution (General Response (1) Problem setting); in the BEV task, for example, the aggregator needs the distribution of the fleet's state of charge, because the state of charge at grid connection and the departure time vary across vehicles. Channel-independent forecasters such as PatchTST share weights across series and predict each series separately, so they provide no cross-vehicle covariance, whereas the variance of the fleet total, 1⊤Σ1\mathbf{1}^\top\Sigma\mathbf{1}, depends on it. Forecasters that attend across series, such as iTransformer and Crossformer, still output point forecasts without a predictive covariance. In contrast, MV-GPSSM outputs this covariance, and its row–column structure makes it estimable from few weekly transitions (General Response (4) Row–column structure). On the two datasets used, we compare MV-GPSSM with baselines that provide joint predictive distributions, and it attains the lowest NLL on both: on the BEV data at large fleet sizes with N=20N=20 (Table 1) and N=10N=10 training weeks (Table R1 in General Response (4)), and on the weather data at every size (Table 10). The BEV data are proprietary, and the public weather data are provided for reproducibility (General Response (2) Datasets).

Reviewer_QD1nOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses spatiotemporal modeling and one-step-ahead forecasting of matrix-valued time series. The authors formulated a matrix-variate Gaussian process state-space model (MV-GPSSM) with structured latent dynamics. They analyzed the computational complexity of the latent-space MVGP inference, explicitly accounting for the costs of lifting and maximum a posteriori (MAP) filtering. Experiments on two empirical datasets (battery electric vehicle fleet data and weather data) show that the MV-GPSSM maintains predictive performance comparable to that of classical baseline methods.

Critical strengths

The paper is clearly written and well organized. The mathematical formulations are presented clearly. Detailed derivations and supplementary materials further enhance the readability and technical clarity of the paper.

Critical weaknesses
  1. The paper feels like one at least 5 years ago. The problem that they worked on is matrix-valued time series, while the attention nowadays is mainly high-dimensional tensor-valued time series. The time series task is the easiest: only one-step-ahead prediction, while nowadays the attention is long-term prediction. The techniques are quite standard and the derivation has possible severe Math issues.

  2. The method only tested on two empirical datasets and only the second one has reproducible code submitted. It lacks thorough examination through simulation studies of the properties of the proposed method. Further it was not compared with any modern AI techniques in this problem setting.

Questions for authors
  1. In the paper, standard GP regression conditioning is used to obtain P(G_{t+1}|G_t,\mathcal{D}{tr}). The inputs {G_1,...,G{N-1}} and {G_2,...,G_{N}} are from the same dynamics. In an ordinary GP regression, the inputs are treated as fixed. However, the inputs are generated recursively by the same random transition function in a GPSSM. Would conditioning on the full latent trajectory provide additional information about the transition function beyond what is captured by the standard GP regression likelihood?

  2. Do the connections to PARATUCK2 and BTD Type-2 produce genuinely new restrictions and properties? The current results that multiplicative latent interactions induce Kronecker-product covariance and additive interactions induce Kronecker-sum covariance, feel like following directly from standard Gaussian linear transformations and independence

  3. Can you provide thorough examination of the proposed method through extensive simulation studies? It is a standard practice in time series literature. Only testing on two empirical datasets is rather weak on experiments.

  4. Can you compare your methods to various modern AI methods, such as Crossformer, iTransformer, and PatchTST?

  5. What is the performance of the proposed method on multiple-steps-ahead forecasting?

3 comments
AuthorsOpenReview ↗

Response to Reviewer QD1n

We thank the reviewer for the detailed questions.

Q1. Would conditioning on the full latent trajectory provide additional information about the transition function?

For learning, no. Although the inputs are generated recursively, the transition function FF enters the density of the latent trajectory only through its values at G1,…,GN−1G_1,\dots,G_{N-1}. By the chain rule, p(G2:N∣G1)=∏tp(Gt+1∣G1:t)p(G_{2:N}\mid G_1)=\prod_t p(G_{t+1}\mid G_{1:t}), where each factor is the GP predictive conditioned on the preceding transition pairs, and this product equals the joint GP marginal likelihood N(y;0,K)\mathcal{N}(y;0,K) with KK in Eq. (13), maximized in our M-step (Frigola et al., 2013, Eqs. (3)–(4)). We have revised Appendix E.1 (shown in red) to write this term as p(G2:N∣G1)p(G_{2:N}\mid G_1) instead of p(G2:N∣G1:N−1)p(G_{2:N}\mid G_{1:N-1}), which suggested fixed inputs. In prediction, the transitions in the prediction trajectory also carry information about FF; we condition the GP on Dtr\mathcal{D}_{\mathrm{tr}} only and do not update it with them, which makes the latent process Markov. We have revised Section 5.3 and Appendices D.3 and E.3 (shown in red) to state this as an approximation instead of attributing the Markov property to the model itself.

Q2. Do the connections to PARATUCK2 and BTD Type-2 produce genuinely new restrictions and properties?

We agree that the Kronecker forms follow from standard Gaussian identities, and we do not claim new constraints from the decompositions. What we add is the correspondence: the latent interaction mechanism (multiplicative or additive) determines the covariance class (Kronecker product or sum). A new latent-level synthetic study (Appendix G.3.6) shows that this correspondence is recoverable from data: from 19 transitions, as in our data, selecting the kernel by validation NLL recovers the generating mechanism in 10/10 seeds for both purely multiplicative and purely additive dynamics.

Q3. Can you provide extensive simulation studies?

We agree that simulation is valuable, and we have added a latent-level simulation (the synthetic study in our answer to Q2; Appendix G.3.6). In this simulation, we generate latent matrix sequences directly, with transitions built from random tanh features following the additive (Eq. (4)) or multiplicative (Eq. (3)) dynamics, and train each kernel on 19 transitions, as in our real data. The kernel selected by validation NLL recovers the generating mechanism in 10/10 seeds for both dynamics. The simulation covers only the latent dynamics; the observation model and the scaling in MM are evaluated on real data. The BEV data are operational records of up to 800 vehicles, the very regime the model targets (Figure 1, Table 1), and the public weather data are provided with code for reproducibility (General Response (2) Datasets).

Q4. Can you compare with Crossformer, iTransformer, and PatchTST?

Please see General Response (5) Deep forecasters; we add the statistical view here. Our target is the joint predictive distribution (General Response (1) Problem setting); in the BEV task, for example, the aggregator needs it because the state of charge at grid connection and the departure time vary across vehicles. Our evaluation is the full-covariance NLL, a proper scoring rule for the joint predictive distribution, so the relevant comparators are models that define such a distribution. Crossformer, iTransformer, and PatchTST are trained with an MSE loss for point forecasting and do not define a predictive covariance. The baselines therefore all define one: GP state-space models that differ from ours only in the row–column structure (Vec. Indep-, ICM-, and LMC-GPSSM), MAR and a dynamic matrix factor model (Matrix SSM) from the matrix time-series literature, and VAR as a vectorized baseline.

Q5. What is the performance on multi-step-ahead forecasting?

In our formulation, each matrix XkX_k collects a whole period, so one step in kk is already a direct multi-step forecast of all LL time points in the period, with their joint covariance; in the BEV task, this is a one-week-ahead forecast of 42 four-hour samples, so Table 1 already reports multi-step performance. The span and resolution of the columns are design choices set by the horizon over which the joint distribution is needed, so a longer horizon (e.g., four weeks) can in principle be forecast directly with a longer span rather than by recursion. This direct design avoids recursive prediction, which accumulates errors and requires propagating uncertainty through the nonlinear GP transition at high computational cost (Girard et al., NIPS 2002). For vehicle SOC, Takano and Jimbo (2025) found that recursive autoregressive prediction and day-by-day GP prediction did not outperform the weekly matrix formulation.

Thank you for the clarification, but I remain unconvinced by the responses to Q4 and Q5 at least.

Q4. The comparison does not need to be restricted to methods that have exactly the same probabilistic output as yours. Crossformer, iTransformer, PatchTST, etc. can still be compared using RMSE or other point-forecast metrics already reported in the paper. Requiring all baselines to provide a full predictive covariance effectively excludes modern forecasting methods by construction and does not establish empirical competitiveness.

Q5. I do not agree that predicting 42 entries inside Xt+1X_{t+1} constitutes multi-step-ahead forecasting in the relevant methodological sense. Regardless of how XtX_t is defined, or how many time points are packed into one matrix, the proposed model still performs the transition Xt⟶Xt+1X_t \longrightarrow X_{t+1}, which is one-step forecasting at the model's time scale.

Simply enlarging the column dimension so that Xt+1X_{t+1} contains four weeks does not demonstrate a methodological ability to perform multi-step dynamical prediction. What is needed is evidence that the proposed methodology can predict Xt+hX_{t+h} for h>1h>1, and, more importantly, that the claimed methodological innovation provides an advantage as the forecasting horizon increases. Otherwise, the response mainly changes the definition of a ``step'' rather than demonstrating genuine multi-step forecasting capability.

Therefore, I still consider at least Q4 and Q5 unresolved.

AuthorsOpenReview ↗

Response to Reviewer QD1n (Q4 and Q5)

We thank the reviewer for the follow-up, and respond to both points below. On Q5 we also clarify one statement of ours that invited a misreading.

Q4. We agree that these methods can be compared using RMSE, which is already used as one of the metrics in the paper, even though they do not provide a joint predictive distribution. The absence of comparisons with modern forecasters limits the conclusions we can draw about point-forecast competitiveness. We note that the paper claims point accuracy comparable to the baselines we evaluated, not superiority, and that its main claim concerns the joint predictive distribution. We will state this limitation of the comparison scope explicitly.

Q5. We clarify one sentence of our earlier answer. What sets the column dimension L is not the horizon one would like, but the periodicity present in the data: L is chosen so that one matrix covers exactly one cycle. This makes the cycle-to-cycle transition approximately time-invariant and allows a single transition map to be learned from the few transitions available. If the data contain a longer cycle, L is chosen to cover it. In either case it is one period; we do not enlarge L in order to lengthen the horizon. Recursive iteration is not part of the intended operation, as our earlier answer also stated. We will state this design rationale in the paper. Even so, the reviewer’s point stands. All reported experiments advance the transition once (Appendix G.1), the paper claims no multi-step prediction that advances the transition more than once, and no result depends on it. The model is, by design, a one-cycle-ahead forecaster, and we state that horizon as an explicit limitation.

Reviewer_kKSxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the Matrix-Variate Gaussian Process State-Space Model (MV-GPSSM), a scalable probabilistic framework designed for matrix-valued time series that preserves coupled row, column, and temporal dependencies. The model maps high-dimensional observations to low-dimensional latent matrices using Tucker2-based representations (incorporating PARATUCK2-type multiplicative and BTD Type-2 additive interactions), models their temporal evolution via a matrix-variate GP, and projects them back through learned bases. A key theoretical contribution is demonstrating that structured matrix-valued kernels arise naturally as consequences of these assumed latent interactions, providing a dynamical interpretation of the covariances.

Critical strengths

This paper proposes a matrix-variate Gaussian process state-space model (MV-GPSSM), which can preserve the row, column and temporal structure while being tractable for high-dimensional data. It is based on well-established tensor decomposition and MVGP, which provides good property to the model.

Critical weaknesses

The paper exploits the well-known MVGP to model the latent factors in a tensor decomposition model, which seems to be incremental. And the experiments in the main content is not extensive, and does show its superior performance.

Questions for authors
  1. The matrix G is already a latent variable that incorporate time information in the Tucker decomposition. Why bother decompose it using PARATUCK2?
  2. The idea is simple, i.e., decompose the matrix variable tensor data using tensor decomposition and model the latent matrix using MVGP. The EM algorithm is also well-established. So the contribution seems to be incremental.
  3. There are many existing works that consider the integration of time information in tensor decomposition, e.g., Bayesian Continuous-Time Tucker Decomposition.
  4. From Table 1, the proposed method has close performance with the baseline methods. The only advantage seems to be the computational complexity.
  5. I am not sure the kernel structure in Fig 2 can provide valid interpretability analysis.
4 comments
AuthorsOpenReview ↗

Response to Reviewer kKSx (1/2)

We thank the reviewer for the careful reading and constructive questions. The revisions are shown in red in the revised PDF and listed in General Response (6) Revisions.

Q1. Why decompose GG further with PARATUCK2?

We decompose GtG_t further because this is what allows us to structure the latent dynamics. Tucker2 alone leaves GtG_t unconstrained and says nothing about how GtG_t evolves to Gt+1G_{t+1}. PARATUCK2 writes Gt=diag(αt)Hdiag(βt)G_t=\mathrm{diag}(\alpha_t) H \mathrm{diag}(\beta_t), separating the row-mode and column-mode contributions so that the dynamics of each mode can be modeled; letting them depend on the previous state gives the transition map F(G)=ΦR(G)WΦC(G)⊤F(G)=\Phi_{\mathrm{R}}(G) W \Phi_{\mathrm{C}}(G)^\top (Eq. (3)), and a matrix-normal prior on WW yields the Kronecker-product kernel (Eq. (8)). BTD Type-2 gives the Kronecker-sum case. The two structures are distinguishable from data (Appendix G.3.6). Without this decomposition, the transition would need a generic kernel on vec(Gt)\mathrm{vec}(G_t), which is the unstructured setting that becomes overconfident (Vec. LMC-GPSSM, Table 1). In Table 1, this structure is what separates MV-GPSSM from Vec. LMC-GPSSM, which uses a GP transition without it.

Q2. Is the contribution incremental?

The tensor decompositions and the MVGP we build on are established, as you note. What the paper adds is the link between them, also identified in your summary: the decomposition of the latent interactions in GtG_t determines the transition kernel, PARATUCK2 giving the Kronecker product and BTD Type-2 the Kronecker sum (see our answer to Q1). The Kronecker structure is thus derived from a decomposition model rather than assumed for tractability, as is common in MVGPs, and each kernel has a dynamical interpretation as an interaction mechanism. In particular, the Kronecker sum is non-separable and lies outside the matrix-normal family underlying the standard MVGP, and it has rarely been used as a GP kernel (e.g., Lan, JMLR 2022, for spatio-temporal observations). Moreover, a new latent-level synthetic study (Appendix G.3.6) shows that this correspondence between mechanism and kernel is recoverable from data. In our setting (General Response (1) Problem setting), this structure is what keeps the joint distribution from becoming severely overconfident with few weekly transitions, even when the training weeks are halved to N=10N=10 (Table 1 and Table R1; General Response (4) Row–column structure).

AuthorsOpenReview ↗

Response to Reviewer kKSx (2/2)

(Continued; Q3–Q5.)

Q3. Related work on time-aware tensor decompositions

Thank you; we have added this line of work to Related Work (Section 2). Time-aware tensor decompositions follow two lines.

  • Time as an index. Bayesian Continuous-Time Tucker Decomposition (BCTT) places a GP over tt on each core element (Fang et al., 2022) and is evaluated by completing missing entries within the observed period. The core trajectory learned in this way does not carry over to a new trajectory, whereas in our setting the prediction trajectory Ht\mathcal{H}_t differs from the training trajectory Dtr\mathcal{D}_{\mathrm{tr}} (Section 3).
  • Transition on latent states. Dynamic matrix factor models, discussed in our Related Work (Yuan et al., 2023; Yu et al., 2024), place a parametric bilinear transition on a latent matrix, and multilinear dynamical systems place a multilinear one on a latent tensor (Rogers et al., 2013). Our Matrix SSM baseline follows Yu et al. (2024): Gt+1=RGtC⊤+VtG_{t+1}=R G_t C^\top+V_t. MV-GPSSM instead models a nonlinear transition Gt+1=F(Gt)G_{t+1}=F(G_t) with a GP that accounts for row- and column-wise covariances, so prediction is conditioned on the current state and gives a joint predictive distribution.
  • Difference from BCTT. In the matrix case, both share the Tucker2 observation model X=AGB⊤X=A G B^\top and differ in where the GP is placed:
    • BCTT: t↦G(t)t\mapsto G(t), an independent GP over time for each element gpq(t)g_{pq}(t);
    • MV-GPSSM: Gt↦Gt+1G_t\mapsto G_{t+1}, a GP whose covariance couples the core elements through the row–column structure.
  • Numbers (Table 1; Table R1 in the General Response): at N=20N=20, the full-covariance NLL is −0.158-0.158 for MV-GPSSM, 2.52×1042.52\times10^4 for Matrix SSM, and 2.32×1042.32\times10^4 for Vec. Indep-GPSSM, which keeps the element-wise independent GPs like BCTT but takes the state as input. At N=10N=10, MV-GPSSM attains 0.720.72, whereas both baselines remain of order 10410^4 (Table R1).
  • References added to Related Work
    • Shikai Fang, Akil Narayan, Robert M. Kirby, and Shandian Zhe. Bayesian Continuous-Time Tucker Decomposition. ICML, 2022.
    • Mark Rogers, Lei Li, and Stuart Russell. Multilinear Dynamical Systems for Tensor Time Series. NeurIPS, 2013.

Q4. Is the advantage only computational?

The advantage is not only computational. Our abstract described both RMSE and NLL as “comparable”, which summarized the results too coarsely. Table 1 and the full results (Tables 7 and 8) show comparable RMSE, but the lowest NLL among the compared methods that remain tractable at large sizes, orders of magnitude lower than the latent-space GP baselines at every fleet size, and these baselines are severely overconfident (normalized Mahalanobis statistic, added in Appendix G.3.1). This is the relevant result whenever the joint predictive distribution is the target (General Response (1) Problem setting); in the BEV task, for example, the aggregator needs the distribution of the fleet's state of charge, which varies with each vehicle's state of charge at grid connection and departure time. The same pattern holds on the weather data (Table 10). We have revised the abstract to state RMSE and NLL separately (General Response (3) Claims).

Q5. Does Fig. 2 provide valid interpretability?

We agree that Fig. 2 is a qualitative check, not a validation, and the text states this (the last sentence of Section 6.2). Its purpose is to show that the learned row and column covariances are consistent with known structure (spatial relations among cities and multi-day persistence), not to claim new physical findings. In the BEV task, the row covariance represents cross-vehicle dependence, which enters the variance of the fleet total.

Thank you for the detailed responses and revisions. The additional clarifications and experiments have helped address some of my concerns and strengthened the empirical evidence. However, I still have reservations about the degree of methodological novelty, as the main theoretical constructions rely largely on established mathematical formulations. Therefore, I will keep my score unchanged.

AuthorsOpenReview ↗

Response to Reviewer kKSx

We thank the reviewer for the careful reading and for the time spent on the discussion, and we are glad that the clarifications and the added analyses were useful. On novelty, our position is unchanged from the General Response. The algebraic identities we use are standard. What we claim as new is the correspondence itself — a multiplicative row–column interaction in the latent dynamics yields a Kronecker-product kernel, an additive one a Kronecker-sum kernel — together with the finding that the generating mechanism can be recovered from as few transitions as our data provide.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers (1/2)

We thank all reviewers. We summarize the scope of the paper and the revisions.

(1) Problem setting. MV-GPSSM targets matrix-valued time series in which the joint predictive distribution over rows and columns is required and the state dimension MLML can grow while the number of transitions NN stays small. Our motivating application is the use of battery electric vehicles (BEVs) in virtual power plants (VPPs) as a source of flexibility for renewable energy. Because a BEV's battery serves driving first, the state of charge at grid connection and the time of departure vary across vehicles, so the charge and discharge capacity of the aggregated fleet is inherently uncertain. An aggregator therefore needs the predicted distribution of the state of charge of the connected vehicles over the coming week, jointly over the fleet, rather than a point forecast. We take one week as the forecasting unit because passenger-vehicle use follows a weekly cycle of weekdays and weekends. One transition is obtained per week of operation; vehicles can be added, but weeks cannot be accelerated, so the fleet's own history provides far fewer transitions than the state dimension MLML (MM vehicles × LL time points per week), i.e., N−1≪MLN-1 \ll ML. The weather data lie in the same regime (N−1=40≪ML=2,100N-1=40 \ll ML=2{,}100 at M=50M=50).

(2) Datasets. The BEV data in (1) are proprietary operational records. The public weather dataset, with executable code, is provided for reproducibility. We have revised Section 6.2 and Appendix G.3.2 to describe this experiment as a reproducibility check of the proposed structure.

(3) Claims. Point accuracy (RMSE) is comparable to the strongest GP baselines, while MV-GPSSM attains the lowest full-covariance NLL among the compared methods feasible at large scale, and, with the latent dimensions fixed at P=10P=10, Q=18Q=18, NLL comparable to the observation-space MVGP at the small sizes where the latter is feasible. We revised the abstract and conclusion to state this precisely.

(4) Row–column structure. The central point of the paper is the row–column structure of the GP transition, which plays two roles in the regime of (1). First, it links the interaction mechanism to the kernel: entry (p,q)(p,q) of Gt+1G_{t+1} depends on GtG_t only through its pp-th row and qq-th column, and multiplicative interaction yields the Kronecker product and additive interaction the Kronecker sum. The algebraic identities involved are standard; what is new is this correspondence and its recoverability, shown by a synthetic study added in the revision (Appendix G.3.6): from 19 transitions as in our data, validation NLL selects the generating mechanism. Second, it makes the joint covariance estimable from few transitions: after the Tucker2 decomposition reduces the state from MLML to PQPQ dimensions, the row–column structure leaves O(P2+Q2)\mathcal{O}(P^2 + Q^2) covariance entries instead of O(P2Q2)\mathcal{O}(P^2 Q^2), with O(P+Q)\mathcal{O}(P + Q) RBF hyperparameters, independent of the fleet size MM (Figure 1). Vec. Indep- and Vec. LMC-GPSSM also use GP transitions and differ only in this structure, so the gap to them is the effect of the structure: halving the training weeks from N=20N=20 (Table 1, ML=33,600ML=33{,}600 at M=800M=800) to N=10N=10 (Table R1 below), MV-GPSSM keeps NLL of order one and a normalized Mahalanobis statistic of about 3–4 (ideal value 1; Appendix G.3.1), whereas theirs remain of order 10310^3–10410^4; RMSE is nearly equal (0.230–0.248), so the methods differ in the predictive distribution, not the point forecast.

Table R1. Results at M=800M=800 with N=10N=10 and N=20N=20 training weeks (same test weeks). For N=20N=20, NLL and RMSE are those reported in Table 1 of the paper; MD²/(ML) is the normalized Mahalanobis statistic in Table 9 (Appendix G.3.1). NLL: full-covariance NLL per element. MD²/(ML): squared Mahalanobis distance of the forecast error divided by MLML (ideal value 1). Mean (std) over 5 seeds.

MethodNLL (N=10N=10)NLL (N=20N=20, Table 1)MD²/(ML) (N=10N=10)MD²/(ML) (N=20N=20)RMSE (N=10N=10)RMSE (N=20N=20, Table 1)
MV-GPSSM (ours)0.723 (0.342)−0.158 (0.012)3.86 (0.57)3.04 (0.02)0.231 (0.005)0.230 (0.001)
Vec. Indep-GPSSM2.74 (0.90) × 10⁴2.32 (0.69) × 10⁴5.48 (1.80) × 10⁴4.64 (1.39) × 10⁴0.248 (0.010)0.239 (0.002)
Vec. LMC-GPSSM2.83 (2.50) × 10⁴4.28 (3.98) × 10³5.67 (5.01) × 10⁴8.56 (7.96) × 10³0.245 (0.013)0.238 (0.002)
Matrix SSM2.56 (0.20) × 10⁴2.52 (0.17) × 10⁴5.13 (0.41) × 10⁴5.04 (0.34) × 10⁴0.263 (0.009)0.256 (0.003)
AuthorsOpenReview ↗

General Response to All Reviewers (2/2)

(5) Deep forecasters. PatchTST, iTransformer, and Crossformer are point forecasters trained with an MSE loss and do not provide an ML×MLML\times ML predictive covariance. Our claim concerns the effect of the vehicle × time row–column structure in the GP transition, so we compare against models that also use a GP transition and differ only in this structure (Vec. Indep- and Vec. LMC-GPSSM), leaving a comparison with general-purpose forecasters for future work. Given the setting in (1) and the evaluation in (4), our baselines are methods that provide joint predictive distributions and can be estimated from few transitions: Vec. Indep/ICM/LMC-GPSSM, Matrix SSM, MAR, and VAR.

(6) Revisions (shown in red in the revised PDF). The model and the main results are unchanged; the revisions clarify the text, add related work, and add two appendix analyses (Appendices G.3.1 and G.3.6).

  • Abstract and conclusion: RMSE and NLL stated separately (Reviewer kKSx).
  • Section 2 (Related work): time-aware tensor decomposition (Reviewer kKSx).
  • Section 5.3 and Appendices D.3, E.1, and E.3: notation of the transition likelihood, conditioning on Dtr\mathcal{D}_{\mathrm{tr}}, and the fixed transition used in prediction stated as an approximation (Reviewer QD1n).
  • Section 6.2 and Appendix G.3.2: the weather experiment described as a reproducibility check (Reviewer QD1n).
  • Appendix G.3.1: calibration diagnostic added (normalized Mahalanobis statistic, Table 9) (Reviewer kKSx).
  • Appendix G.3.6 (new): a latent-level synthetic study showing that selecting between the Kronecker-sum and Kronecker-product kernels by validation NLL recovers the generating interaction mechanism (Reviewers QD1n and kKSx).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.