- 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- Outdated related work and evaluation.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]).