- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes EMBER, an operator for graph-indexed spatio-temporal forecasting whose central object is a coupled boundary-to-future map: the observed history and the graph jointly define a "coordinated boundary state", and all forecast horizons are generated directly from that state rather than by autoregressive rollout.
Validation is on six highway datasets (PeMS03/04/07/08, PeMSD7(M)/(L)), two urban-flow datasets (T-Drive, CHIBike, inflow/outflow), and a univariate Weather2K-R cross-domain experiment. The paper also proves two projection/decomposition propositions, a "path dependence and reversal barrier" theorem, a power-law accumulation theorem, and two corollaries, and reports three mechanism diagnostics.
Critical strengths
1. The problem framing is genuinely well-motivated, and one component is a real contribution.
Most spatio-temporal forecasting work assumes the future is a smooth continuation of the recent past, and most STGNNs symmetrize the graph before propagation. This paper takes seriously that (a) relations are directional and signed, (b) observed dynamics can rupture rather than decay, and (c) the same normalized response can correspond to different physical changes at different nodes. The "robust boundary fiber" plus the scale-restoration at readout (Eq. 2 and Eq. 15, with Corollary 2 giving a conditional positive-affine equivariance statement) is the most defensible and most transferable piece of the design: using median/MAD rather than mean/std, and explicitly multiplying back the per-node robust scale, is a principled treatment of the non-stationarity that plagues these benchmarks. This is the part of the paper I would expect others to reuse.
2. Accuracy is competitive across a broad and heterogeneous set of tasks.
EMBER has the best MAE on 10 of 10 traffic tasks and the best RMSE on 6 of 10, spanning sensor graphs and grids, 12-step and 1-step protocols, and speed and flow targets, with a positive cross-domain signal on three meteorological variables. Breadth of this kind is not trivial to achieve, and it is the main reason I am not recommending a clear rejection.
Critical weaknesses
W1. The empirical evidence does not support the paper's causal claims: standard errors are absent, margins are within the baseline spread, and two of the four pillars are never properly ablated.
This is the decisive weakness.
(a) No variance, no seeds. Tables 2–4 report single numbers with no standard deviation, no confidence interval, and no statement of how many runs were averaged. This matters concretely here, because the winning margins are of the same order as the spread among the top baselines. On PeMSD7(M), EMBER's MAE is 2.57 against STGNCDE's 2.68 — a 4.1% gap, while five other baselines (STGODE 2.97, MSTGCN 3.54, ASTGCN 3.61, DCRNN 3.83, STGCN 4.01) are all far worse than that, but the top five methods in that column span only 0.31 MAE. On PeMS03, the gap to TFormer is 3.4%; on PEMS08, 3.4% to STAEformer; on PeMS07, 2.7% to RAST. In the traffic-forecasting literature these gaps are routinely smaller than seed-to-seed variation on identical code. Without std over at least three seeds, it is not possible to tell whether EMBER is better than STAEformer/RAST/PDFormer at all, and every headline claim in the abstract ("strongest overall performance") rests on that.
(b) The paper's own diagnostics undercut two of its four pillars. Figure 2(c) reports the learned fractional order as having "mean 0.094 and range 0.087–0.138" across 3,567 validation windows. A standard deviation of roughly 0.012 on a quantity that is described throughout — in the abstract ("sample-level fractional memory law"), in Section 3.7 ("sample-conditioned fractional order"), and in the conclusion ("long-memory aggregation") — as sample-conditioned is essentially a constant. Worse, for the kernel gives , , , so the readout places roughly 78% of its total weight on the single most recent step (total mass at ), versus 8.3% under uniform weighting. The learned operator is therefore close to a one-step readout with a thin tail, not a "power-law long-memory aggregation". Note also that sits near the hard floor imposed by the parameterization in Eq. 23, which suggests the model is actively pushing the fractional mechanism toward the identity. The same issue applies to the "Galerkin chart": Figure 2(b) shows the retained subspace captures a projected energy ratio of only ~0.44–0.47, i.e. less than half the latent variation, which the text nevertheless describes as "substantial". And Figure 2(a) is circular as a diagnostic: Proposition 1 proves the whitened Gram matrix equals exactly, so the reported "after" condition number of is guaranteed by construction and is not evidence of anything.
Crucially, no ablation isolates these claims. "w/o Fractional" is vacuous on T-Drive Inflow (the authors' own footnote: makes the operator an identity), so the fractional readout is ablated on exactly one dataset, where it moves MAE by 0.22. There is no variant with a single learned scalar , and no variant with linear (uniform) accumulation — precisely the comparisons that would determine whether the "fractional" story is doing work.
(c) Direct multi-horizon decoding is never ablated, despite being Challenge II in the introduction and one of the four contribution bullets. Every competing method in Tables 2–3 also forecasts all horizons, so the paper's central contrast against autoregressive decoding is never tested against an autoregressive version of EMBER itself.
(d) The ablations confound mechanism removal with capacity and input removal. "w/o Galerkin" deletes an entire residual block (Eq. 22), not just the whitening or the low-rank projection; the reported degradation of 0.65 MAE therefore mixes "the Galerkin mechanism matters" with "one fewer block matters". "w/o Boundary Fiber" removes an input pathway, changing the input dimensionality. Both need capacity-matched replacements (e.g., a standard MLP block of equal width/depth; raw windowed features instead of fiber statistics).
W2. The theoretical results are properties of the chosen algebraic form, not statements about learning; they cannot adjudicate the design choices, and one of them is quantitatively vacuous at the horizons tested.
Propositions 1–2 and Theorem 1–2 are proved correctly — I checked the proofs in Appendix B and the unrolling and Gamma-ratio arguments are fine. But their content is definitional:
- Prop. 1 says an orthogonal projector is an orthogonal projector with spectral norm 1 and that a full-column-rank whitening gives Gram and condition number 1. True, and true of any whitening.
- Prop. 2 says an orthonormal frame plus its orthogonal complement reconstructs a vector losslessly. This is the Pythagorean theorem, applied to a frame the model was constrained to construct via modified Gram–Schmidt (Eq. 8). The "exactness" is enforced by the parameterization, not learned.
- Theorem 1's reversal barrier is a statement about the accumulator , and the proof shows only that two different preceding sums yield two different growth values — which is immediate from being a sum. The bound follows from and monotonicity of tanh. It quantifies the arithmetic of a running mean, not any behavior of the trained network.
- Theorem 2 states that generalized-binomial weights decay and that their partial sums are . This is the standard asymptotic of the fractional-difference kernel; it is a property of the kernel the authors chose, not a discovered property of the model. Moreover, with the measured and (and for the urban tasks, where the mechanism is the identity by the paper's own footnote), ; the "sublinear accumulation" is barely distinguishable from linear over the horizons actually evaluated.
The abstract promises that the analysis "establishes stable Galerkin coordination, exact rupture decomposition, a quantifiable evidence barrier ..., and positive power-law memory". In each case the property is built into the functional form by construction. None of the four results distinguishes EMBER from a simpler model with the same constraints, and none connects to the experiments. Together with W1(b), the theory and the diagnostics are largely restating the architecture back to the reader.
W3. Specific result overclaiming, including one contradiction with the authors' own table.
Section 5.4 states that "Tables 2 and 3 show consistent predictive gains" and the abstract claims "the strongest overall performance among the compared methods". However, in Table 3 EMBER is not best on three of four urban-flow RMSE columns:
| Setting | EMBER RMSE | Best baseline RMSE |
|---|---|---|
| T-Drive Outflow | 30.78 | 30.61 (PDFormer) |
| CHIBike Inflow | 5.34 | 5.31 (PDFormer) |
| CHIBike Outflow | 5.29 | 5.27 (STAEformer) |
The text carefully says "lowest MAE on all four" — which is true — but the surrounding claim of consistent gains, and the RMSE framing in PEMS03 ("the 2.46-point reduction ... is particularly clear", which is an RMSE figure positioned immediately after a list of MAE figures), create an impression of uniform superiority that the table does not support. On the six highway datasets, the text also says EMBER "obtains the lowest MAE and RMSE in every dataset", while PEMS04 MAE is a 18.06/18.06 tie with STAEformer (acknowledged in the same sentence, then contradicted by it).
W4. Baseline comparability is unspecified, and the cross-domain experiment uses a materially weaker baseline pool.
The paper does not state whether baseline numbers were re-run under EMBER's splits and preprocessing or copied from the original papers. This is not a pedantic point for these benchmarks: PeMS03/04/07/08 have documented inconsistencies in split ratios and evaluation masks across the literature, and TFormer/PDFormer/STAEformer numbers differ by several MAE points depending on protocol. Without a per-baseline statement, the margins in W1(a) cannot be attributed to the method.
Relatedly, the Weather2K-R comparison (Appendix D, Table 6) uses DCRNN, STGCN, MSTGCN, TGCN, and CLCLR — none of the strong baselines from Table 2 (STAEformer, PDFormer, RAST, STGNCDE, MFormer). The claimed "cross-domain transfer" is therefore confounded with a weaker comparison set, and the univariate, distance-adjacency setup is considerably easier than the multivariate traffic setting it is being used to support.
Questions for authors
- Variance. How many seeds were run for Tables 2–4? Please report mean ± std over at least three seeds, or state explicitly that the numbers are single runs. Given the 2.7–4.2% MAE margins and the spread among top baselines, this determines whether the results support any of the paper's comparative claims.
- Baseline protocol. For each baseline in Tables 2, 3, and 6, were numbers re-run under your splits, preprocessing, and mask, or copied from the original papers? Please add a table stating this per method and per dataset (especially split ratios for PeMS03/04/07/08).
- Table 3 RMSE. EMBER's RMSE is second or third on T-Drive Outflow, CHIBike Inflow, and CHIBike Outflow. How do you reconcile this with "consistent predictive gains" in §5.4 and "strongest overall performance" in the abstract?
- Fractional order. Figure 2(c) gives a mean of 0.094 and a std of about 0.012, near the parameterization floor. What does the boundary-conditioned buy over a single learned scalar (or over uniform linear accumulation)? Please provide both ablations, on a dataset with .
- Direct multi-horizon decoding. Please provide an ablation in which EMBER's readout is replaced by autoregressive rollout. This is Challenge II in the introduction and is never tested.
- Rupture validity. No diagnostic shows the rupture/hysteresis mechanism triggering at real events. Can you show, on PEMS08 or T-Drive, that the accumulated charge and the modal growth coefficient change sign or magnitude at congestion onset and clearance, as the motivation in §1 predicts? A synthetic system with known regime switches would also serve.
- Weather baselines. Why is the Weather2K-R comparison restricted to four older models plus CLCLR, omitting every strong method used in Table 2?
- Capacity-matched ablations. Can "w/o Galerkin" and "w/o Boundary Fiber" be replaced with capacity-matched and input-matched controls (a standard residual block of equal width; raw windowed features instead of fiber statistics)?
- Please provide , learning rate, batch size, epochs, parameter counts, and training/inference wall-clock relative to STAEformer on PEMS08, and release code.
Dear Reviewer,
Thank you again for your valuable comments and feedback. As the rebuttal deadline is approaching, we would like to kindly follow up on our responses, in which we have addressed your questions and concerns.
We would greatly appreciate it if you could revisit our responses and consider whether they warrant an upward revision of your score. If you have any remaining questions or concerns, please do not hesitate to let us know. We would be more than happy to provide further clarification.
Thank you for your time and consideration.
Best regards,
The Authors