acceptodds
Under review as a conference paper at ICLR 2027

FORECASTING AS CONDITIONAL TRANSPORT: DOOB- REVERSIBLE GIBBS DYNAMICS ON SPATIO-TEMPORAL GRAPHS

Abstract

Observed variation in a spatio-temporal graph mixes history-predictable evolution with conditional surprise. Their distinction depends on the same available history, motivating coordinates that expose both the learned predictable reference and its residual. We frame forecasting around the construction and transport of predictive coordinates: represent variation relative to available history, propagate the resulting evidence through graph-informed interactions, and retain access to it across forecast horizons. DR-GIBBSFORMER realizes this view with a causal multi-offset Student-t projector, dual-stream temporal–spatial Gibbs attention, triangular exchange, and horizon-specific full-history functionals. The analysis characterizes coordinate bias and residual-score behavior, content–graph factorization, operator sensitivity, recoverability under exchange, and linear readout expressivity. Across ten highway and urban-flow tasks, the framework improves 18 of 20 aggregate MAE/RMSE comparisons, with gains reaching 8.26% in MAE and 7.61% in RMSE. Component and projector ablations support the value of learned conditional coordinates, structural interaction, and access to the terminal history. These findings connect the choice of predictive coordinates to the way historical evidence is exchanged and used for forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Time Series Forecasting, Traffic Forecasting, Conditional Forecasting, Spatio-Temporal Graph Learning, Dynamical Systems, Gibbs Attention

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

This paper proposes DR-GibbsFormer, a spatio-temporal graph forecasting model. The method first builds history-conditioned "level" and "innovation" coordinates with a causal Student-t projector. It then processes them with temporal and spatial attention biased by a directed graph prior (called Gibbs transport), exchanges information between the two streams with an additive coupling layer, and predicts each horizon with a linear readout over the full encoded history. The paper includes several propositions about these components. The method is evaluated on six highway datasets (PeMS) and two urban-flow datasets (T-Drive and CHIBike).

Critical strengths

The method obtains lower average MAE/RMSE than the listed baselines in 18 of 20 comparisons, and the ablations in Tables 3 and 4 show that each component contributes to the final performance.

Critical weaknesses

I recommend rejection for this paper. The paper is difficult to follow, the motivation is not clear, and the experimental evaluation doesn't reflect the current state of the art in spatio-temporal graph forecasting.

  • Presentation and motivation. The abstract and the introduction are very hard to read. The introduction goes directly to technical details ("conditional coordinates", "Gibbs transport", "triangular exchange", "horizon-specific full-history functionals") without explaining the context or the problem being solved. Concepts are introduced out of the blue, and the introduction reads like a list of technical components. The method section has the same problem. As a result, I don't see why Gibbs transport or the Doob decomposition is a good idea for forecasting in the first place. As acknowledged in Section 2, the "Gibbs transport" is scaled dot-product attention with an additive graph bias, and the "reversible" exchange is a standard additive coupling layer. Similarly, most propositions are standard results (conditional expectation, bounded Student-t score, tower property, softmax as an entropy-regularized problem, invertibility of additive coupling, rank factorization), and they don't explain why the method works better. Proposition 1 also doesn't show that the learned residuals are innovations in the Doob sense, since the Student-t location is trained jointly with other losses and doesn't need to be the conditional mean.
  • Outdated related work and evaluation. The related work on spatio-temporal graph forecasting stops in 2023, and it misses the significant progress in recent NeurIPS, ICML, and ICLR papers (e.g., [R1, R2]). The baselines also miss recent methods from these venues. Moreover, the current challenge in spatio-temporal forecasting is performance across multiple horizons. Tables 1 and 2 only report averages, and Figure 2 and Appendix C show per-horizon results only for the proposed method on two datasets, with no comparison against the state of the art. Finally, the baseline numbers have no standard deviations; it's unclear whether they follow the same protocol (e.g., the masking of non-positive targets in Appendix A.3), and some wins are smaller than the reported standard deviation across seeds (e.g., PeMS04 MAE: 18.053 ± 0.082 vs. 18.06).

References

[R1] "Continuous product graph neural networks", NeurIPS 2024.

[R2] "SONAR: Long-Range Graph Propagation Through Information Waves", NeurIPS 2025.

Questions for authors

Clarifications:

  • What's the motivation for framing forecasting as conditional transport? What does "Gibbs transport" offer beyond attention with an additive graph bias?
  • How does the method compare with the baselines at individual horizons (e.g., 3, 6, and 12 steps)?
  • Were the baseline results reproduced under the same splits, normalization, and masking protocol, or copied from the original papers?
  • Do the learned residuals behave as innovations (e.g., conditional mean close to zero and no serial correlation per node)?

Suggestion for the revision:

  • Compare against recent spatio-temporal graph models published at NeurIPS, ICML, and ICLR (e.g., [R1, R2]).
Reviewer_qkStOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DR-GibbsFormer, a spatio-temporal graph forecasting model built around history-conditioned level and innovation representations. A causal Student-t projector estimates conditional reference values and constructs residual-based features, with auxiliary supervision at multiple temporal offsets. The model combines temporal and spatial Gibbs attention, a graph-informed structural prior, and triangular cross-stream exchange. Horizon-specific readout functions access the complete encoded history to predict multiple future steps directly. Experiments on six highway benchmarks and four urban-flow tasks report improvements in 18 of 20 aggregate MAE/RMSE comparisons, supported by component ablations and theoretical analyses.

Critical strengths

The conditional innovation representation is a meaningful modeling choice, and the comparison with first-difference coordinates provides direct empirical evidence for its usefulness.

The evaluation covers ten forecasting tasks across road networks and urban-flow grids, offering broader evidence than a single-dataset evaluation.

The ablation study separately examines the innovation stream, level stream, graph prior, cross-stream exchange, and full-history readout.

The paper provides mathematical analyses of conditional residual bias, Student-t residual scores, exchange invertibility, and readout expressivity.

The implementation details include model dimensions, optimization settings, data-split procedures, and three-seed results, which support reproducibility.

Critical weaknesses
  1. The proposed architecture combines several existing components, and the paper does not sufficiently isolate the methodological novelty of Gibbs attention from standard row-wise softmax attention with a structural bias.

  2. The theoretical results establish several properties of individual components, but the connection between these properties and the overall forecasting improvements remains limited.

  3. The main ablation study covers only PeMS08 and T-Drive Inflow, leaving the component contributions insufficiently verified across the full benchmark suite.

  4. The model introduces multiple streams, a causal GRU projector, and separate temporal and spatial attention operations, but the experimental section does not report inference latency, memory consumption, or a comprehensive efficiency comparison.

  5. The reported improvements are not consistent across all metrics: PeMS04 RMSE is slightly worse than the strongest baseline, while CHIBike Inflow RMSE also underperforms the best listed result. The practical significance of the gains therefore requires further examination.

Questions for authors
  1. How does the proposed Gibbs attention compare with standard scaled dot-product attention using the same graph bias, parameter budget, and training configuration?

  2. Can the authors provide controlled experiments that isolate the effects of conditional coordinate construction, Student-t supervision, and triangular exchange across more than two datasets?

  3. How sensitive are the results to the projector offsets, loss weights, graph-prior mixing coefficient, and model width?

  4. What are the parameter count, inference latency, peak GPU memory, and training cost compared with the strongest baselines under matched hardware and evaluation settings?

  5. Given the mixed RMSE results on PeMS04 and CHIBike Inflow, do the improvements remain statistically reliable across seeds, and how do they vary across individual forecast horizons?

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

The paper introduces DR-GIBBSFORMER, a supervised spatiotemporal graph forecaster built to leverage two key representations of available historical data. A causal GRU component predicts Student-t locations/scales at many different offset levels, and a dual-stream Gibbs encoder uses spatial and temporal attention guided by a directed graph prior. Information exchange between streams is facilitated by a subsequent triangle exchange step, and horizon-specific linear contractions read the complete terminal history, combining with an innovation path and future features. The method is demonstrated on 10 related traffic/urban-flow tasks, and shows improvement on 18 of 20 aggregate MAE/RMSE comparisons.

Critical strengths
  1. The paper's stated contributions are clear and attempt to improve performance in the challenging domain of spatiotemporal forecasting. The introduced approach is sensible, and presents a testable inductive bias, with ablations providing positive evidence for its utility. The positioning of the work is additionally well-motivated, and the core gaps identified in the introduction are key limitations of many existing methods, and addressing those limitations serves as a foundation for novelty with opportunity for high impact.
  2. The full model demonstrates fairly broad performance gains over the application domain, along with appropriately qualified mathematical statements. Tables 1 and 2 report seeded means and standard deviations, and include both average-error gains and exceptions (e.g., w/ CHIBike Inflow improving MAE but performing worse in terms of RMSE). There are additional discussions that provide useful contextual details and qualifications, e.g., Appendix B.1 distinguishes a fitted Student-t location from a true conditional mean under misspecification and limits reversibility to the exchange operation.
Critical weaknesses
  1. I found the core methodological novelty of the proposed approach to be somewhat weak. Section 2 mentions how the one-marginal Gibbs operator is scaled dot-product attention with a structural prior, and attributes the triangular coupling scheme to reversible networks. The core components of the approach are therefore fairly familiar mechanisms, making it somewhat difficult to classify it as a fundamentally distinct approach that brings along new theoretical properties.
  2. The performance advantages on the primary benchmark appear to have some meaningful interpretability qualifications. In Section 4.1, it's mentioned that available benchmark values are used, but it's never clarified which settings reference available values and which were rerun. Further, for those values that were used, it's unclear whether there's a shared tuning and preprocessing pipeline that would allow the results to be directly compared in this manner, and details like data splits/seeding/etc appear to only be provided for PeMS04 and PeMSD7(M) (and not the other scenarios; I could have missed some other setting details but only these two are mentioned in A.3). The comparison concerns here are mainly important due to the several marginal performance wins (e.g., PeMS04 MAE is 18.053 vs 18.06, PeMS07 RMSE is 32.798 vs 32.97, and CHIBike Outflow RMSE is 5.261 vs 5.27). Against unreplicated baselines, specific seeds (for only 3 trials) is not sufficient to measure uncertainty or establish a clear avenue for comparability.
Questions for authors
  1. Can you clarify which entries in Tables 1-2 were rerun and which were used as available values? Details like task-specific split boundaries, sampling intervals, and baseline preprocessing would be useful to report in order better understand how the comparisons across baselines are made. Do the larger performance gains and near ties persist under a shared pipeline and repeated baseline runs?
  2. Can a comparison be provided between fixed settings, e.g., the same graph-biased, full-history backbone with matched capacity, and auxiliary prediction supervision, but a conventional single or generic two-stream representation? Similarly, can you clarify what exactly replaces each removed component in Table 3? This would distinguish the conditional-coordinate contribution from additional parameters and training signals.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.