acceptodds
Under review as a conference paper at ICLR 2027

Spatial-as-Refinement: Gated Graph Operators over Strong Temporal Predictors for Scalable Spatio-Temporal Forecasting

21 views

Abstract

Graph-based spatio-temporal forecasters route information through a learned adjacency at every encoder layer, with graph operators as the substrate and temporal operators interleaved at every depth. We argue this convention is structurally fragile when the learned adjacency is noisy, as it often is on fragmented sensor networks. Under a perturbed-adjacency model, we show that substrate placement yields a noise bound that compounds multiplicatively through the operator norms of all other layers, so the optimizer cannot silence one bad layer without discarding the signal it carries. We propose GraFT, built on a structural reordering we call spatial-as-refinement: a per-sensor temporal predictor first produces a complete forecast, and a small stack of -gated graph blocks refines it through residual updates , each gate initialized to zero. Under the same model, this reordering replaces the multiplicative bound with one that is exactly zero when gates are closed and confines noise to a channel each layer's gate controls independently. The same operator is therefore a liability or an asset depending only on where it sits, and the optimizer decides how much to use it from data alone. Experiments on controlled synthetic and real-world data confirm the theory's three predictions (substrate models diverge under adjacency noise, refinement-structured models degrade gracefully, and the mean gate closes under fragmentation) and let us verify the multiplicative-versus-gated mechanism directly. Across fourteen real-world benchmarks spanning traffic, energy, meteorology, and mobility, a single GraFT instantiation establishes new state-of-the-art performance, attaining the best average rank against strong published baselines, and the learned gates reveal where the graph actually helps.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal forecasting, Graph neural networks, Time-series Analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kPb9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GraFT, a spatio-temporal forecasting architecture based on a “spatial-as-refinement” principle. Instead of interleaving graph operators throughout the forecasting backbone, GraFT first produces a complete per-sensor temporal forecast and then applies several gated graph refinement blocks. The authors argue theoretically that conventional substrate-style graph placement can amplify adjacency noise multiplicatively across layers, whereas zero-initialized refinement gates can suppress graph-induced contamination. The paper evaluates this idea through synthetic experiments and fourteen real-world benchmarks across traffic, energy, meteorology, and mobility, reporting strong forecasting performance and a correlation between learned gate magnitudes and graph connectivity.

Critical strengths

The strongest part of the paper is the breadth of the empirical evaluation combined with a clear architectural hypothesis. GraFT performs strongly across fourteen datasets, including very large networks such as LargeST-CA with 8,600 sensors, and the improvements are not limited to one temporal backbone because MLP, PatchTST, and TSMixer variants are all evaluated with the same refinement stack. I also appreciate that the authors attempt to connect theory and experiments through controlled perturbation experiments rather than presenting the theoretical analysis independently of the empirical section.

Critical weaknesses

The central weakness is that the claimed mechanism is only partially demonstrated. The theory contrasts an ungated substrate with a gated refinement architecture, so the resulting bounds naturally favor the latter when the gates approach zero. But this does not establish that refinement placement itself is responsible for the improvement; a gated substrate may also suppress noisy graph contributions. The authors explicitly state that they do not include this comparison. In addition, Proposition 4 is analyzed for a LayerNorm-free linear refinement path, while the deployed architecture contains LayerNorm and optional FFNs, and the simplified small-gate interpretation does not apply to the trained gate magnitudes of 0.28–1.18. These qualifications do not make the theoretical analysis incorrect, but they substantially narrow what can be concluded from it about the deployed model.

Questions for authors

My main question is whether the central conclusion survives a controlled gated-substrate versus gated-refinement comparison using the same graph operator, gate initialization, predictor capacity, and training protocol. Without this experiment, it is difficult to distinguish the benefit of moving the graph after the temporal predictor from the benefit of simply giving the optimizer an explicit scalar mechanism to suppress the graph. I would also like the authors to clarify how strongly they intend the theoretical noise argument to describe the deployed GraFT architecture, given that the main bound is derived on the LayerNorm-free path and the trained gates lie well outside the stated small-gate regime.

Reviewer_vQDLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work uses a GNN to fit the residuals between the predictions of a temporal model and the ground truth — a simple yet effective idea. The experimental results themselves may be valuable, but the interpretation of the underlying mechanism goes beyond what the current evidence supports. In particular, the paper lacks a gated-substrate ablation, which directly undermines the validity of the core contribution rather than merely making the experiments less comprehensive.

Critical strengths
  1. The separation between per-sensor temporal prediction and optional cross-sensor refinement is conceptually clear. Closing the graph gates preserves a graph-independent forecasting path, allowing the architecture to retain temporal modeling capacity without requiring spatial mixing. The design is also modular: the same refinement stack can be combined with different temporal backbones.

  2. The paper evaluates GraFT on fourteen benchmarks spanning traffic, energy, meteorology, and mobility, including large sensor networks. The use of MLP, PatchTST, and TSMixer backbones provides useful evidence that the approach is not restricted to a single temporal architecture. The reported results are competitive across multiple settings, supporting the practical relevance of the proposed framework.

Critical weaknesses
  1. The central contribution of graph placement is not isolated from gating. The proposed method changes both the placement of graph computation and the use of zero-initialized residual gates. However, the evaluation does not include a substrate architecture equipped with the same gated residual mechanism, as acknowledged in Appendix I. Consequently, the results do not establish whether the gains arise from post-forecast placement, gating, or their interaction. An interleaved architecture of the form $h^{(\ell+1)} T_\ell!\left(h^{(\ell)}+\alpha_\ell G_\ell h^{(\ell)}\right)$ can also preserve a complete per-sensor temporal path when all gates are closed. Thus, this property is not unique to post-forecast refinement. Given that gated residual connections have clear precedents, a matched placement-by-gating comparison is essential to substantiate the claimed architectural novelty. Such a comparison could initially be conducted using the MLP backbone rather than requiring new variants of every backbone.

  2. Lemma 1 bounds the actual factorized graph operator GG. To apply this bound to the ideal operator (G⋆)(G^\star), the authors need to explicitly assume $0\le G^\star\le G$ elementwise, or prove that G⋆G^\star has a structure governed by the same mode-load upper bound MM. Under the additive perturbation assumptions currently stated, one can only guarantee $(|G^\star|_\infty\le M+\epsilon$).

  3. The synthetic experiments do not cleanly identify the proposed adjacency-noise mechanism.

Questions for authors

Please see weaknesses.

Reviewer_4wyaOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Graph-based spatio-temporal forecasters utilize information through a learned adjacency at every encoder layer. The paper points out that the learned adjacency may become noisy hence reduce the performance of graph-based model. This paper proposes GraFT, built on a structural reordering called spatial-as-refinement with gate mechanism. Across fourteen real-world benchmarks, GraFT achieves state-of-the-art performance.

Critical strengths

The paper proposes GraFT, which incorporates a "spatial-as-refinement" mechanism. The synthetic experiment effectively validates the claims regarding the gate component of GraFT. The model achieves state-of-the-art (SOTA) results across multiple datasets.

Critical weaknesses

As a \textbf{graph-based} model, GraFT explicitly claims that it does not require any prior physical adjacency matrix and learns a graph operator through learnable factors such as E1E_1 and E2E_2. However, it is unclear that the learned E1E_1, E2E_2 and wi,j(l)w_{i,j}^{(l)} actually capture \textbf{meaningful graph structures.} The authors should provide a explanation of what the learned graph represents. In particular, could the authors visualize the learned graph matrix WW and compare its connectivity patterns with the ground-truth physical adjacency matrix? I do not necessarily expect the learned graph to be highly similar to the physical graph. However, at least one case study showing the learned graph and discussing its structural differences from the physical graph would be important for understanding what the model actually learns. This is particularly important given the mathematical formulation of the graph operator. Since WW is constructed from the softmax outputs of E1E_1 and E2E_2, its entries are strictly positive. Consequently, the resulting matrix is inherently dense. Without showing that the learned weights exhibit meaningful concentration or distinctive connectivity patterns, it is unclear whether the resulting structure represents a meaningful graph with informative neighborhood relationships, \textit{or simply acts as a generic dense linear mixing operator}.

Questions for authors
  1. The synthetic experiments in Section 4 have provided a useful analysis of the role of the proposed gate mechanism. However, I still have the question about the connection between synthetic experiments and the theoretical analysis. The theoretical analysis is formulated in terms of Gℓ=Gℓ∗+Nℓ\mathcal{G}_\ell=\mathcal{G}_\ell^* + \mathcal{N}_\ell. Although Appendix C.4 directly injects perturbations, it does not show whether the graph operators learned by GraFT actually contain noisy components or how they relate to an informative graph Gℓ∗\mathcal{G}_\ell^*. I think the assumption that the learned graph contains noise is the motivation of the gate mechanism. \textbf{Could the authors compare the learned graph with a reference graph (e.g., one constructed from the known community structure in the synthetic setting) and characterize their difference?} This would provide a more direct connection between the actual graph-learning component and the theoretical perturbation analysis.

  2. In Sec. 5.3, I do not understand the following sentence: `` Fixing the gates at αℓ=1\alpha_\ell=1 is never free either, though its cost tracks connectivity less cleanly than the trained mean gate ∣αℓ∣|\alpha_\ell|, which falls monotonically with connectivity, mirroring the synthetic readout under real, heterogeneous graphs’’. \textbf{Connectivity is not the same as graph quality. How is graph quality measured?}

  3. What are the performances of PatchTST, MLP, and TSMixer? The paper should compare the proposed method with these backbones to demonstrate the improvements brought by the graph-based module.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.