acceptodds
Under review as a conference paper at ICLR 2027

FRAME: RETHINKING SPATIO-TEMPORAL FORECASTING FROM OBSERVED BOUNDARIES TO FUTURE FIELDS

Abstract

Spatio-temporal forecasting asks how an observed relational field should extend into an unknown future. We cast this task as boundary-conditioned extension: history supplies a measured boundary, directed relations define its geometry, and the graph–horizon product domain supports the future field. We introduce FRAME (Future-field Reconstruction through Action-Minimizing Extension), which separates this construction into coordinate-consistent boundary representation, explicit continuation of measured structure, and joint inference of innovation. Robust local affine fibers and sample-conditioned whitened Galerkin projections coordinate the boundary, while a constant-preserving affine trace carries observed level and trajectory forward. Innovation is defined by the unique minimizer of a strongly convex robust action coupling conservative graph exchange, temporal variation, curvature, and boundary attachment, and is approximated by learned solver steps. Our analysis establishes affine factorization, non-expansive projection, uniqueness, and bounded conservative structural influence. Direct numerical verification further shows that every learned update descends the stated action toward its variational target and that four steps closely reproduce the prediction induced by the converged target. Across twelve tasks spanning ten datasets, including ChinaTemp and ChinaWind, FRAME attains the lowest MAE on every task and the lowest RMSE on all six highway and both meteorological benchmarks.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Time Series Forecasting, Dynamical Systems, Multi-Horizon Forecasting, Variational Forecasting, Neural Operators, Graph-Based Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rHM1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reformulates spatio-temporal forecasting on a graph as a boundary-conditioned extension problem. The observed history is treated as a boundary. The H×N future field is obtained in three parts: a per-node, per-sample representation of the history, an explicit continuation of the observed level and local trend, and a residual innovation field that approximately minimizes a convex objective over the graph×horizon product lattice. Concretely, FRAME first normalizes each node's history with a robust affine transform and lifts it, together with node, calendar, and spectral graph-position embeddings, into a D-dimensional state. It then mixes nodes through L blocks of low-rank, QR-whitened Galerkin projection with gated residuals. An affine-constrained attention over history steps produces a continuation value ρ for each horizon. A K-dimensional drive per horizon and node is decoded and refined with S=4 shared, learned preconditioned gradient steps on an objective composed of an L2 fidelity term and log-cosh penalties on graph differences, first and second temporal differences, and a first-horizon boundary term. The prediction is read out as ρ plus two learned corrections, one from the refined field and one from the full-width context, and mapped back through the node's affine fiber.

The theoretical statements are that the normalization is invariant to per-node positive affine changes, the whitened projection is an orthogonal projector, the objective is 1-strongly convex, and the graph flux is bounded and sums to zero across nodes. Validation covers six PeMS highway benchmarks, the T-Drive and CHIBike urban-flow benchmarks, and the ChinaTemp and ChinaWind meteorological benchmarks, against up to 16 baselines per task. It also includes component and action-term ablations, solver-depth sensitivity, and a numerical check that the four unrolled steps descend the stated objective toward an L-BFGS minimizer. FRAME outperforms the baseline algorithms across diverse benchmark datasets.

Critical strengths

S1. The future field has an explicit, verified optimization target. FRAME defines the future innovation as the minimizer of a stated strongly convex objective and unrolls four learned steps toward it. Appendix B.3 checks this against an L-BFGS solve of the same objective: every step descends, and the final prediction differs from the converged one by about 0.1%.

S2. Broad and internally consistent evaluation. Twelve tasks cover traffic flow, speed, urban flow and weather, horizons of 1 to 30 steps, and graphs of 170 to 1,026 nodes. Ablations in Tables 4 and 5 degrade monotonically in the expected order, and the one-step tasks show the gain is not only multi-horizon smoothing.

Critical weaknesses

W1. Baseline numbers are unsourced and inconsistent with the original publications. Table 2 does not say whether baselines were re-run or copied. For STAEformer the listed numbers differ from the original paper, and against the published numbers FRAME's lead shrinks to at most 0.2 MAE on PeMS04, PeMS07 and PeMS08. STGformer, which reports lower PeMS04 MAE than FRAME, is not compared. The claim of lowest MAE on every task is therefore not established.

W2. No cost reporting, and the components are known constructions under new names. No parameter count, training time, latency or memory is reported, although each block runs a per-sample QR and the solver sweeps all edges and horizons four times. The components are robust normalization, SVD graph positional encoding, QR-orthogonalized low-rank attention, sum-to-one attention over history, and a log-cosh smoothness regularizer solved by unrolled gradient descent. Prior work deriving GNN layers from such denoising objectives are not discussed. The theory is correct but elementary. The contribution is a careful combination, so the paper rests on its empirical claim, where W1 applies.

Questions for authors
  1. Can you list all hyperparameters per dataset?
  2. Can you state whether code will be released?
Reviewer_Ty9aOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper formulates spatio-temporal forecasting as extending a future field from observed boundaries and proposes FRAME, which has three steps:

  1. It constructs the boundary representation from a node-level robust affine decomposition and a sample-conditioned, QR-whitened Galerkin projection.
  2. It explicitly continues the observed level and trend with an affine trajectory whose weights sum to 1.
  3. On the graph × horizon lattice, it defines a strongly convex log-cosh robust energy on a reduced-dimension "innovation field", approximately solves it with 4 learned preconditioned gradient steps, and reads out the prediction as "affine continuation + energy correction + full-width context correction".

The main text reports the best MAE on all 6 highway tasks and all 4 urban-flow tasks.

Critical strengths
  1. The problem decomposition (boundary / continuation / innovation) is clear, and the energy (Eqs. (10)–(13)) and the properties stated in Sec. 3.7 are precisely stated and mathematically correct.
  2. The evaluation covers highway traffic (flow and speed) and urban flow, and the reported MAE is the best on all 10 tasks.
  3. The design is simple: the constant-preserving affine trajectory and the robust local normalization are easy to reuse.
Critical weaknesses
  1. No positioning against the closest prior work; some components overlap with existing constructions. The "variational/optimization-based prediction" part of Sec. 2 cites only an algorithm-unrolling survey and OptNet, and none of the following related work is mentioned:
    • The robust affine fiber (Eqs. (2), (16)) "removes level and scale per window and restores them after prediction", which is essentially instance normalization and already exists in RevIN [1] and its spatio-temporal counterpart ST-Norm. Such methods restore the output with input-window statistics, which implicitly assumes that the forecast window and the input window are identically distributed; Dish-TS [2] specifically points out and handles the failure of this assumption. The paper does not discuss this. What happens when the window is nearly constant (the scale reaches ε) or when the trend changes abruptly (e.g., at the start of the morning peak)?
    • RelCoord (Sec. 3.2) is, as described in the text, "directed edge lengths → two-hop affinity → truncated SVD → source/target-side coordinates", which is very close in spirit to HOPE [3]. What is the difference?
  2. The "variational extension" claim does not match what the model actually computes and constrains.
    • Theorem 1 only guarantees that, for fixed (D, B, gates), the exact minimizer exists and is unique. It is equivalent to a proximal map applied to D and uses only "a strongly convex quadratic term plus a convex potential". What the model actually computes is S=4 iterations in which the step size η and the preconditioner Π are both learned (Eq. (13)). The main text does not relate η and Π to the Lipschitz constant determined by the learned κ and β, and gives no error bound for A^(S) approximating A*; "approximated by a finite-step solver" is a statement, not a proof.
    • End-to-end SPEN [4] backpropagates the error through iterative gradient-based inference, so the energy is tuned for the prediction quality of the approximate inference. This means that joint training learns the behavior of the truncated inference, which need not coincide with the exact minimizer, so the unique-minimizer property constrains the model output only weakly. The numerical verification mentioned in the abstract is also not in the main text.
    • Propositions 1–3 and the elementary properties in Theorem 1 follow from zero-sum conservative exchange, which holds for any graph divergence-form term, so they are not specific to this energy. None of these results concerns the S-step solver that is actually used, nor the final prediction.
  3. The comparison experiments could be more comprehensive.
    • The baseline set omits strong baselines that are widely compared on the same datasets: GWNet [5], STID, and D2STGNN. The original papers of STID and D2STGNN report PeMS04/08. AGCRN is cited in Sec. 2 but does not appear in Tables 2/3.
  4. The core selling point (joint coordination across horizons) cannot be tested with the main-text metrics, and the unfavorable evidence on urban flow is not discussed.
    • Table 2 reports only the 12-step average MAE/RMSE. The standard practice in Graph WaveNet [5] is to report MAE/RMSE/MAPE at the 15/30/60-minute horizons. Without per-horizon results, one cannot tell whether the improvement comes from long horizons (where coordination should help) or short horizons. The urban-flow tasks have H=1, so the temporal terms vanish by construction and do not test this either.
Questions for authors
  1. Could the authors add GWNet, STID, and D2STGNN (and discuss STD-MAE), as well as per-horizon (15/30/60-minute) results and MAPE on the PeMS tasks?
  2. Relative to RevIN/ST-Norm/Dish-TS, HOPE, SPEN and end-to-end SPEN, graph trend filtering, and Elastic GNN/TWIRLS, what exactly does FRAME add?

References

[1] Kim, Kim, Tae, Park, Choi, Choo. Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift. ICLR 2022.

[2] Fan, Wang, Wang, Wang, Zhou, Fu. Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting. AAAI 2023.

[3] Ou, Cui, Pei, Zhang, Zhu. Asymmetric Transitivity Preserving Graph Embedding. KDD 2016.

[4] Belanger, Yang, McCallum. End-to-End Learning for Structured Prediction Energy Networks. ICML 2017.

[5] Wu, Pan, Long, Jiang, Zhang. Graph WaveNet for Deep Spatial-Temporal Graph Modeling. IJCAI 2019.

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

the paper proposes FRAME for forecasting spatial temporal graphs. It combines a input window normalization, low-rank projection, and a weighted average based trend prediction with correction and a gradient based smoothing. It shows strong MAE and RMSE results on the benchmarks listed.

Critical strengths

None

Critical weaknesses

The paper is very poor written, filled with terminologies that were never properly defined or explained, while the underlying operations behind these terminologies were established and had well-understood names (like "robust local affine boundary fibers" which is essentially normalization). Writing issues are all over the place: like Gc□PH (formatting issue, the operation is likely not a compilable symbole), extremely sparse-written related works, and practically no qualitative or insightful understanding of the numbers. I would not recommend accepting this paper.

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.