acceptodds
Under review as a conference paper at ICLR 2027

Spatio-Temporal Decomposition: A Plug-in Interpretable Front End for Multivariate Time Series Forecasting

Abstract

Forecasting models for spatio-temporal series usually either ignore cross-location structure or encode it with a graph that is fixed in advance and hard to read. We propose Spatio-Temporal Decomposition (STD), a module that sits in front of any forecasting backbone and splits each observation into four named components, namely local and global in space and high and low frequency in time. The split uses only means over time and over locations plus two sets of learned offsets, one per location and one per time step, so it introduces no additional hyperparameters beyond those of the backbone. We show that the per-location offset converges to the correction a piecewise-stationary trend requires, and that the per-time offset acts as an implicit set of spatial weights, so the plain average over locations becomes a learned neighborhood without a graph. Across TCN, Transformer and WaveNet backbones on BikeNYC, SKT, and Traffic, STD lowers MAE in 26 of 27 backbone, dataset and horizon combinations, cuts MAE summed over the three horizons by 19 to 29% on BikeNYC, and lifts graph-free backbones above MTGNN and StemGNN. On BikeNYC the global high-frequency part shows two commuting peaks on weekdays and one on weekends, indicating that STD surfaces readable structure that graph-based models encode only implicitly. Ablation experiments consistently show that removing spatial aggregation increases mean RMSE, highlighting the contribution of spatial information to forecasting accuracy. Robustness evaluations on Capital further show that STD maintains stable performance under Gaussian noise and achieves the lowest MAE and RMSE among competing models at 20% missingness.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal forecasting, Multivariate time series, Interpretable deep learning, Plug-in module

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_V39kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Spatio-Temporal Decomposition (STD), a plug-in, interpretable front-end module for multivariate time series forecasting. STD circumvents the need for predefined, static spatial graphs by decomposing each observation into four interpretable components: local and global in space, and high and low frequency in time. This is achieved using simple temporal and spatial means along with two sets of learned offsets. The authors provide theoretical guarantees demonstrating that these learned offsets can implicitly adapt to piecewise-stationary trends and capture spatial interaction weights without explicit graph structures. Extensive experiments across diverse backbones (TCN, Transformer, WaveNet) and real-world datasets (BikeNYC, Traffic, SKT, Capital) show that STD consistently improves forecasting accuracy while offering qualitative interpretability into the data's structural dynamics

Critical strengths
  1. The plug-in nature of STD is commendable. It does not introduce complex hyperparameter tuning but relies on simple mean aggregations and learnable offsets to expand the input into a 9-dimensional descriptive feature vector.

  2. This work substantiates its architecture with rigorous theory. Theorem 1 (Trend Adaptation), Theorem 2 (Implicit Weight Learning), and Theorem 3 (Optimization Convergence) successfully bridge the gap between simple offset learning and complex spatiotemporal dynamics.

  3. By proving that the global time-specific offset ϕt\phi_t acts as an implicit set of spatial weights, the paper offers a computationally cheap yet effective alternative to memory-heavy Graph Neural Networks requiring explicit adjacency matrices.

  4. Evaluated against robust backbones (TCN, Transformer, WaveNet) on diverse datasets, STD demonstrates pervasive improvements and shows resilience to data corruption.

Critical weaknesses
  1. The main comparison tables primarily feature slightly dated models like LSTM, GRU, LSTNet, MTGNN, and StemGNN. While recent SOTA models (like iTransformer and PatchTST) are mentioned in the extended appendices, they should be central to the main manuscript to definitively prove STD's modern utility.

  2. STD increases the input feature dimension by a factor of 9 (Zjt∈R9Z_{jt} \in \mathbb{R}^9). As explicitly shown in Table 13, this leads to an inevitable increase in the number of parameters for the input layer of the backbones, which increases training time.

  3. Theorem 2 theoretically claims that ϕt\phi_t approximates true spatial interaction weights. However, there is no empirical visualization or quantitative comparison showing how these learned implicit weights correlate with actual geographic distances or ground-truth adjacency matrices.

  4. The global spatial representation relies on a simple spatial mean. For extremely large systems with thousands of highly heterogeneous sensors, this simplistic mean might wash out granular spatial features before the offset ϕt\phi_t can correct them.

  5. Some recent plug-and-play methods for multivariate time series forecasting are not discussed in the paper, such as: [1] Wang Z, Jin G. Don’t Just Extract, Regulate: Contextual Residual Exchange for Non-stationary Multivariate Stock Forecasting[C]//Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2026: 253-270.

Questions for authors
  1. Provide a small-scale empirical analysis (e.g., on BikeNYC or PeMS) correlating the implicit spatial weights derived from ϕt\phi_t with the actual spatial geographic distances or physical adjacency matrices. Do the implicitly learned spatial dependencies actually align with geographical reality?

  2. Address the dimension expansion overhead explicitly. Can you provide an ablation study showing if a subset of the 9 channels (e.g., just top 4 components) provides a better accuracy-efficiency Pareto frontier?

  3. Briefly discuss or test the framework's performance limit on spatial datasets with substantially larger node counts (N>1000N > 1000) where global means might suffer from over-smoothing.

Reviewer_W8KCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Spatio-Temporal Decomposition (STD), a plug-in framework that decomposes multivariate time series into local/global spatial and high-/low-frequency temporal components using mean-based operations and learnable offsets. The authors provide theoretical analyses of trend adaptation, implicit spatial weighting, and optimization convergence. Experiments with TCN, Transformer, and WaveNet backbones report improvements on several short-horizon forecasting benchmarks, supplemented by synthetic experiments, interpretability analyses, ablation studies, and robustness evaluations.

Critical strengths
  1. Modular and interpretable design: STD provides a relatively simple, plug-in decomposition framework that explicitly separates spatial and temporal components, offering a potentially useful way to improve forecasting interpretability without redesigning existing backbones.

  2. Promising short-horizon performance: The method demonstrates consistent improvements across TCN, Transformer, and WaveNet on the three primary benchmarks, suggesting practical benefits in short-term spatio-temporal forecasting.

Critical weaknesses

Critical Weaknesses

  1. Fundamentally questionable theoretical validity: The paper's theoretical claims contain serious mathematical weaknesses. Theorem 2 does not establish that a shared scalar offset can recover location-specific spatial interaction weights, while Theorem 3 incorrectly claims that its step-size schedule satisfies the Robbins–Monro conditions. Moreover, subtracting temporal means does not constitute a genuine high-/low-frequency decomposition, and the resulting components lack identifiability guarantees. These issues substantially undermine the paper's central theoretical and interpretability claims.

  2. Limited novelty and insufficient empirical generalization: The proposed framework largely repackages elementary mean-based decomposition and learnable offsets, with substantial conceptual overlap with existing spatio-temporal normalization and structured decomposition methods (Deng et al., 2021; 2024). More critically, Appendix Table 17 shows that STD worsens MSE in six of seven extended-horizon experiments and MAE in all seven. These results directly challenge the claimed general applicability. The reliance on relatively outdated primary baselines further weakens the experimental contribution. Overall, the methodological novelty and technical rigor fall substantially short of ICLR standards.

References

[1] Deng, J., Chen, X., Jiang, R., Song, X., & Tsang, I. (2021). ST-Norm: Spatial and Temporal Normalization for Multi-variate Time Series Forecasting. Proceedings of ACM SIGKDD, 269–278. https://doi.org/10.1145/3447548.3467330

[2] Deng, J., Chen, X., Jiang, R., Yin, D., Yang, Y., Song, X., & Tsang, I. (2024). Disentangling Structured Components: Towards Adaptive, Interpretable and Scalable Time Series Forecasting. IEEE Transactions on Knowledge and Data Engineering, 36(8), 3783–3800. https://doi.org/10.1109/TKDE.2024.3371931

[3] Robbins, H., & Monro, S. (1951). A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22(3), 400–407.

[4] Ghadimi, S., & Lan, G. (2013). Stochastic First- and Zeroth-Order Methods for Nonconvex Stochastic Programming. SIAM Journal on Optimization, 23(4), 2341–2368.

Questions for authors
  1. Theoretical validity: How can Theorem 2 justify recovering location-specific spatial interaction weights from a shared scalar offset? Can the authors establish identifiability and explain whether the theoretical guarantees apply to the learned spatial pooling used in the actual implementation?

  2. Convergence analysis: Theorem 3 incorrectly claims that the proposed step-size schedule satisfies the Robbins–Monro conditions. Furthermore, the claimed gradient Lipschitz bound and multiplicative variance reduction appear insufficiently justified. Can the authors provide mathematically rigorous corrections to these results?

  3. Generalizability and empirical validity: Appendix Table 17 shows that STD worsens MSE in six of seven extended-horizon experiments and MAE in all seven. How do the authors reconcile these findings with their general applicability claims? What evidence demonstrates that the improvements arise from meaningful spatio-temporal decomposition rather than additional representation capacity?

Reviewer_jiDXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Spatio-Temporal Decomposition (STD), a plug-in module for multivariate time series forecasting. STD decomposes time series into local and global spatial components and high- and low-frequency temporal components using mean operations and learnable offsets. The module can be integrated with different forecasting backbones, including TCN, Transformer, and WaveNet. Experiments on multiple datasets demonstrate improvements in forecasting accuracy, while component visualizations and ablation studies support the effectiveness and interpretability of the proposed approach.

Critical strengths

S1. The problem of improving both forecasting accuracy and interpretability in multivariate time series is important.

S2. The plug-in design allows the method to be integrated with different forecasting backbones.

Critical weaknesses

W1. The proposed method mainly combines existing temporal and spatial decomposition operations. Its methodological contribution beyond prior work is unclear.

W2. These sections provide limited discussion of existing methods and fail to clearly motivate the research problem. Despite the available page budget, the discussion remains rather superficial.

W3. The selected baseline models are not sufficiently representative of existing methods for spatio temporal forecasting, particularly recent decomposition-based approaches such as ST-Norm [1].

[1] ST-Norm: Spatial and Temporal Normalization for Multi-variate Time Series Forecasting,KDD 2021

Questions for authors

Q1. What is the difference between the proposed STD and ST-Norm [1]?

Q2. See the weaknesses above.

Reviewer_4yNJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a spatiotemporal decomposition (STD) module designed to integrate seamlessly with various forecasting architectures. The core idea is to decompose spatiotemporal observations into local/global spatial components and high/low temporal frequencies. The authors demonstrate that adding this module improves performance across a diverse range of spatiotemporal baselines, including Transformer-, TCN-, and GNN-based models. Across several real-world datasets, the paper highlights the benefits of STD, particularly for short-horizon forecasting tasks.

Critical strengths

S1. The proposed spatiotemporal decomposition module can be easily integrated into existing spatiotemporal architectures, serving as a general-purpose enhancement for forecasting models.

S2. Short-term forecasting experiments show that STD consistently reduces prediction error across most evaluated baselines.

Critical weaknesses

W1. The manuscript's writing and framing could be improved. The Introduction lacks sufficient motivation for the proposed decomposition, and the Related Work section does not adequately contextualize the method within the broader literature on spatiotemporal forecasting.

W2. The primary empirical evaluation focuses mainly on short prediction horizons, whereas recent literature emphasizes longer forecasting windows. While Table 18 in the Appendix includes preliminary long-horizon experiments, it remains unclear whether the proposed method provides meaningful improvements for larger horizons (e.g., H = 96).

Questions for authors

Q1. Do the performance trends observed in Tables 1 and 2 hold for longer prediction horizons?

Q2. How does the STD module handle exogenous input variables that are not direct targets of prediction?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.