acceptodds
Under review as a conference paper at ICLR 2027

AR-MGNet: Adaptive Macro-Micro Reconciliation Multi-Grid Network for General Spatio-Temporal Forecasting

Abstract

Spatio-temporal forecasting underpins numerous real-world applications, yet existing methods struggle to jointly achieve global-scale efficiency and local-scale fidelity, particularly when system states shift or abrupt events occur. Most approaches rely on fixed spatial partitions or uniform refinement strategies, limiting their ability to adapt to dynamically changing functional regions. To overcome these limitations, we propose AR-MGNet, an Adaptive Macro-Micro Reconciliation Multi-Grid Network that reformulates forecasting as a bidirectional macro-micro reconciliation process. AR-MGNet first aggregates fine-grained observations into functional macro-grids to efficiently solve a coherent macroscopic future, then selectively recovers unresolved micro-scale innovations in event-critical regions through a top-down adaptive refinement pathway. A dynamic multi-grid neighborhood learning module further captures scale-aware spatial dependencies by constructing sparse, relation-aware adjacency matrices that adapt to the current spatial regime. Across four LargeST traffic benchmarks and five diverse general spatio-temporal datasets (wind, air quality, EV charging, electricity, solar), AR-MGNet achieves the best performance in all traffic and most cross-domain datasets, with ablation studies confirming the contribution of each core mechanism. These results demonstrate that adaptive macro-micro reconciliation offers a unified principle for general spatio-temporal forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-temporal forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EispOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes the AR-MGNet network, which predicts multivariate spatiotemporal sequences using a sparse, window-dependent hierarchical structure. It predicts explicit coarse-scale future states, transfers these predictions to finer grids, and applies budget-constrained local residual corrections. Evaluations cover four LargeST regions and five non-traffic datasets, and include additional analyses regarding time series, ablation, and event windows.

Critical strengths

S1. The scope of the assessment is quite broad. By addressing non-transportation improvements in the power, urban electric vehicle (UrbanEV), and solar energy sectors, the study demonstrates the method's practical value beyond a single application scenario.

S2. It is written clearly and is easy to understand.

Critical weaknesses

W1. The evidence for the specific adaptive mechanism is unconvincing and insufficient to support the full model. Although Figure 3 omits certain macro-modules, it fails to distinguish between the "window-dependent routing mechanism" and a "static hierarchical structure with identical prediction heads and loss functions," nor does it differentiate between "learned event selection" and "random selection under the same refinement budget." Thus, these results do not clarify whether the performance gains stem from the adaptive allocation itself or from the additional hierarchical prediction and supervision mechanisms.

W2. Compared to HieraMix, the average MAE improved by only 0.9%, with a difference of 0.11 in MAE for both GBA and CA. Given the marginal performance advantage in traffic forecasting and the model's high sensitivity to depth, it is difficult to assess the reliability of the broad claim that the method "consistently outperforms other approaches." Furthermore, the reported improvements in traffic forecasting accuracy should ideally be supplemented by uncertainty assessments and a clearer explanation regarding sensitivity to the training process.

Questions for authors

Q1. Is there a controlled comparison between dynamic and static allocation when the prediction head, auxiliary loss, and training budget are identical? Similarly, how do learned active masks perform compared to random selection at the same refinement ratio?

Q2. What is the reason for the significant drop in CA performance across the four levels shown in Table 4—optimization, model scale, routing, or other modified settings?

9 comments
AuthorsOpenReview ↗

Author Responses (1/3)

Thank you very much for your insightful and constructive feedback on our submission. The following are our detailed responses.


W1

“The ablations needs to isolate the performance gains of the adaptive routing and learned selection from the additional hierarchical prediction and supervision mechanisms.”

Response:

Thank you for raising this point. To separate the contribution of adaptive routing and learned event selection from hierarchical prediction heads and consistency supervision, we conducted the exact controlled comparisons you requested.

1. Dynamic Routing vs. Static Hierarchical Routing, Isolating Hierarchical Prediction & Supervision Components

Rows 1–5 of following Table present the controlled comparisons for routing and selection. All variants share identical 3-level hierarchy, prediction heads, primal-dual consistency losses, 100% refinement quota, and training schedule. Comparing Row 2 (Static + Learned) against Row 5 (Dynamic + Learned) demonstrates that window-dependent routing yields consistent MAE reductions on CA (16.01 → 15.44), GLA (16.58 → 16.24), SDWPF (0.2260 → 0.2243), and Electricity (0.2026 → 0.1990).

Meanwhile, we hold dynamic routing and learned selection fixed while removing individual hierarchical prediction and supervision modules (Rows 6–9 of following Table). The Table shows that hierarchical supervision (Macro Solver, Consistency Constraints) and adaptive mechanisms (Micro Reconciliation, Fine-level Learned Correction) work in concert: hierarchical supervision anchors the global trajectory, while adaptive mechanisms reorganize regions and concentrate fine-scale computation on high-error nodes.

Controlled Routing/Selection Comparisons and Hierarchical Component Ablations

VariantCAGLASDWPFElectricity
MAERMSEMAERMSEMAERMSEMAERMSE
Static Allocation + Random Selection16.1827.8116.7228.690.22680.39710.20410.3362
Static Allocation + Learned Selection (w/o Dyn. Grid)16.0127.5416.5828.460.22600.39580.20260.3338
Dynamic Routing + Random Selection15.7827.2116.4528.290.22560.39510.20150.3324
Dynamic Routing + Historical-Change Selection15.6026.9816.3928.180.22500.39410.20040.3308
AR-MGNet (Dynamic Routing + Learned Selection)15.4426.7616.2427.950.22430.39290.19900.3286
w/o Macro Solver16.4228.1217.5129.560.22710.39760.20840.3412
w/o Micro Reconciliation15.8627.3416.6828.540.22580.39540.20310.3345
w/o Consistency Constraints16.1527.7816.8928.820.22760.39820.20680.3389
w/o Fine-level Learned Correction16.2827.9517.1229.140.23120.40110.21160.3458

2. Learned Event Selection vs. Equal-Budget Random & Historical-Change Selection

At same 50% refinement ratio, learned selection consistently outperforms both random and historical-change masks across all four benchmarks. Specifically, the post-prediction oracle audit in following Table (mean ± std over 3 seeds) shows that our learned active mask achieves 58.4%–66.2% Top-Error Oracle Precision and captures 62.1%–71.4% of pre-correction inherited errors, substantially outperforming random selection (50.0%) and historical-change selection (56.4%–61.8%). The results are shown as follows.

Post-Prediction Oracle Selection-Quality Audit at 50% Refinement Budget

DatasetActive-Mask SelectorMAERMSETop-Error Oracle PrecisionInherited Error Captured
CALearned Event Router (Ours)16.68 ± 0.0627.82 ± 0.1264.8% ± 0.9%69.7% ± 1.1%
Historical-Change Magnitude16.96 ± 0.0728.15 ± 0.1357.9% ± 0.8%61.8% ± 0.9%
Random Selection17.28 ± 0.1028.52 ± 0.1650.0% ± 0.1%50.0% ± 0.1%
GLALearned Event Router (Ours)17.12 ± 0.0828.86 ± 0.1266.2% ± 0.8%71.4% ± 1.0%
Historical-Change Magnitude17.38 ± 0.0929.18 ± 0.1457.4% ± 0.9%61.2% ± 0.8%
Random Selection17.65 ± 0.1129.51 ± 0.1750.0% ± 0.1%50.0% ± 0.1%
SDWPFLearned Event Router (Ours)0.2258 ± 0.00120.3962 ± 0.001758.4% ± 1.0%62.1% ± 1.0%
Historical-Change Magnitude0.2269 ± 0.00140.3981 ± 0.001953.6% ± 1.4%56.4% ± 1.1%
Random Selection0.2279 ± 0.00210.3998 ± 0.002850.0% ± 0.1%50.0% ± 0.1%
ElectricityLearned Event Router (Ours)0.2142 ± 0.00110.3468 ± 0.001663.5% ± 0.9%68.2% ± 1.1%
Historical-Change Magnitude0.2176 ± 0.00120.3512 ± 0.001856.8% ± 1.0%60.4% ± 0.9%
Random Selection0.2215 ± 0.00170.3564 ± 0.002349.9% ± 0.2%49.9% ± 0.2%
AuthorsOpenReview ↗

Author Responses (2/3)

W2

“Compared to HieraMix, it is difficult to assess the reliability of the broad claim that the method ‘consistently outperforms other approaches.’ Furthermore, the reported improvements in traffic forecasting accuracy should ideally be supplemented by uncertainty assessments and a clearer explanation regarding sensitivity to the training process.”

Response:

Thank you for raising this important point. We have revised the manuscript text to replace broad phrasing with precise, empirically calibrated claims, and supplemented our evaluation with 3-seed uncertainty assessments, paired temporal-block confidence intervals, multi-domain and regime-shift comparisons.

1. AR-MGNet Delivers Substantial Gains Over HieraMix, especially Non-traffic Generalization Prediction Gains

On standard 12-step traffic benchmarks, point MAE across predominantly steady-state windows is highly saturated. AR-MGNet delivers substantial improvements over HieraMix across the following operational dimensions:

  • On Tail & Percentage Errors (RMSE / MAPE), AR-MGNet lowers GBA RMSE by 0.80 (2.6%) and MAPE by 1.76 percentage points (13.8% relative reduction, 10.95% vs. 12.71%), while reducing four-region average MAPE by 5.7%.

  • In Long-Horizon Rollout Stability under the unified evaluation, at h = 48 steps on GLA, AR-MGNet lowers Endpoint MAE by 46.2% (26.55 vs. 49.33) and Cumulative MAE by 37.3% (22.07 vs. 35.19) relative to HieraMix.

  • In Inference Efficiency, AR-MGNet cuts full-test inference time by 46.2% on GBA (14.2s vs. 26.4s) and 56.1% on GLA (17.0s vs. 38.7s, 2.28 times faster) relative to HieraMix.

  • Whereas HieraMix is tailored specifically for road-traffic hierarchies, AR-MGNet is designed for general spatio-temporal forecasting. Across the five non-traffic domains, AR-MGNet reduces MAE by an average of 14.3% relative to HieraMix (26.4% on Solar-Energy, 18.3% on Electricity, 16.0% on UrbanEV, 6.7% on KnowAir, and 3.9% on SDWPF), and achieves lower MAE than the best-performing baseline on 4 of the 5 datasets, with reductions ranging from 6.7% to 19.0%. The results are shown as follows.

Cross-Domain Comparison with HieraMix and the Strongest Baselines

DomainAR-MGNet (Ours)HieraMixBest baseline
MAERMSEMAERMSEMAERMSE
Solar-Energy0.17090.36340.23220.45540.2110 (NEST)0.4347 (TimeMixer)
Electricity0.19900.32860.24370.39090.2295 (PatchSTG)0.3699 (PatchSTG)
UrbanEV4.3610.805.1911.244.89 (STID)11.24 (HieraMix)
KnowAir14.8422.8915.9024.3815.90 (HieraMix)24.38 (HieraMix)

2. Uncertainty Assessments & Structural Resolution of Depth Sensitivity

As reported in Appendix E.1, Table 9, and Table 10 of the revised paper, across 3 independent random seeds (Seeds 42, 43, 44), AR-MGNet's sample standard deviations are 0.05–0.08 MAE on traffic benchmarks (< 0.5% of the mean). Even on GBA and CA where the point MAE margin over HieraMix is 0.11, the 2,000-replicate paired moving-block bootstrap 95% confidence intervals are strictly negative ([-0.18, -0.05] on GBA and [-0.19, -0.04] on CA, p < 0.05).

AuthorsOpenReview ↗

Author Responses (3/3)

Q1

“Is there a controlled comparison between dynamic and static allocation when the prediction head, auxiliary loss, and training budget are identical? Similarly, how do learned active masks perform compared to random selection at the same refinement ratio?”

Response:

Thank you for raising this point. We conducted both controlled experiments under strictly identical prediction heads, auxiliary losses, trainable parameter counts, refinement ratios, and training budgets (see Tables 1 and 2 in W1).

1. Controlled Dynamic vs. Static Allocation

In the dynamic router, affinities are computed from the window-conditioned state and temporal change. For the Static Allocation control, we replace these window-dependent inputs with fixed, training-only node embeddings, passing them through the exact same trainable projection matrices and capacity-balancing iterations. Thus, the parameter count, heads, and losses remain strictly identical.

As shown in Table 1 (W1), Dynamic Routing consistently and significantly outperforms Static Allocation across all four datasets:

  • CA: 16.01 (Static) → 15.44 (Dynamic), a 3.56% reduction (95% CI: [-0.76, -0.38]).
  • GLA: 16.58 (Static) → 16.24 (Dynamic), a 2.05% reduction (95% CI: [-0.48, -0.20]).
  • SDWPF: 0.2260 (Static) → 0.2243 (Dynamic), a 0.75% reduction (95% CI: [-0.0029, -0.0005]).
  • Electricity: 0.2026 (Static) → 0.1990 (Dynamic), a 1.78% reduction (95% CI: [-0.0051, -0.0017]).

2. Learned Active Masks vs. Equal-Budget Random & Historical-Change Selection

To isolate the selection mechanism, we replace the learned event router with a uniform Random mask or a Historical-Change Magnitude mask under identical refinement budgets.

  • At 100% Budget (Table 1 of W1): Learned selection lowers CA MAE to 15.44, outperforming Random (15.78) and Historical-Change (15.60).
  • At 50% Budget (Table 2 of W1): Learned selection lowers CA MAE to 16.68, outperforming Random (17.28) and Historical-Change (16.96). Crucially, it achieves the highest Post-Prediction Oracle Precision (58.4%–66.2% vs. 50.0% for Random and 53.6%–57.9% for Historical Change), proving it precisely targets the nodes with the largest errors.

Q2

“What is the reason for the significant drop in CA performance across the four levels shown in Table 4—optimization, model scale, routing, or other modified settings?”

Response:

We thank the reviewer for this insightful comment. We systematically analyzed all candidate factors mentioned in your question (spatial partitioning structure, optimization convergence, model scale, and routing behavior) and confirmed that the performance drop at L = 4 on CA in original Table 4 (15.44 at L = 3 vs. 23.44 at L = 4) is caused by structural spatial over-coarsening (information bottleneck) under the standard partitioning schedule rather than optimization or routing failure.

  • Primary Structural Cause — Spatial Over-Coarsening at L = 4 (N3=68N_3 = 68 Macro-Cells): In the original Table 4 spatial partitioning scheme (8 → 4 → 4), the 3-level CA hierarchy compressed 8,600 sensors to 8,600 → 1,075 → 269, whereas adding a fourth level compressed the coarsest grid down to only N3=68N_3 = 68 macro-cells. Representing 8,600 statewide California freeway sensors with only 68 macro-cells (averaging more than 126 sensors per cell) creates a severe spatial information bottleneck, merging opposing traffic directions and distinct urban corridors into the same macro state and over-smoothing the inherited forecast Yˉ(l)\bar{\mathbf{Y}}^{(l)}.

  • Optimization & Routing Collapse Ruled Out: Adding the fourth level increases model parameters from 2,423,435 to 2,926,367, and normalized training MAE decreases smoothly during optimization. For the original standard L=4L=4 partition (8,600 → 1,075 → 269 → 68), router diagnostics confirm that the router maintains 96.8%–97.2% parent-cell utilization across the three transitions (1,040.5 / 1,075, 260.8 / 269, and 66.1 / 68), balanced candidate entropy (0.412–0.531), a negligible near-tie fraction (0.008–0.012), and 100% non-zero router gradients, ruling out routing collapse or vanishing gradients.

Effect of the Number of Grid Levels of AR-MGNet on CA

CA settingGrid sizesMAERMSE
L = 1860035.2555.28
L = 28600 → 107522.9835.86
L = 38600 → 1075 → 26915.4426.76
L = 4 (Original Table 4)8600 → 1075 → 269 → 6823.4437.30
L = 4 (Another Option)8600 → 2150 → 538 → 13517.8228.19

We truly appreciate the care and clarity in your feedback. We hope the additional analyses and comparison have fully addressed your remaining concerns. We’ve genuinely learned from your perspective and would be glad to discuss any points that may remain.

Reviewer_xnTyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AR-MGNet for spatiotemporal forecasting. It generates macro-level predictions through bottom-up sparse aggregation, then corrects local residuals using a top-down event selection mechanism, with dynamic multiscale graphs and consistency constraints incorporated into joint training. Experiments cover four traffic regions and five non-traffic datasets, achieving the best results on all traffic datasets and four non-traffic datasets.

Critical strengths
  1. Separately handling macro-level trend prediction and local residual correction helps account for both large-scale spatial dependencies and abrupt local changes, with relatively clear connections between the modules.
  2. In addition to traffic, the evaluation covers wind power, air quality, charging demand, electricity, and solar energy, supporting the architecture’s applicability to different types of spatiotemporal data.
  3. Under the same hardware and batch size, inference time on GLA decreases from 38.7 seconds for HieraMix to 17.0 seconds, while achieving lower prediction errors, demonstrating practical value.
Critical weaknesses
  1. The paper lacks controlled comparisons that incrementally introduce dynamic routing and selective correction under the same backbone, parameter count, and training settings, making it difficult to determine the specific sources of the gains.
  2. Budget sensitivity analysis and a small number of event visualizations are insufficient to demonstrate that the model actually prioritizes the nodes most in need of correction. Comparisons with random selection and selection based on the magnitude of historical changes under the same budget are missing.
  3. In Table 1, NEST’s 12-step MAE on GLA is 28.38, whereas its 16-step MAE in Table 8 decreases to 23.63; HieraMix’s MAE increases from 16.45 to 33.14. The metric definitions, test windows, and implementation details need to be clarified; otherwise, it is difficult to assess the reliability of the reported long-horizon advantage.
Questions for authors

See Weaknesses

5 comments
AuthorsOpenReview ↗

Author Responses (1/3)

We thank the reviewer for the constructive suggestions that helped strengthen the paper. We have addressed each point carefully in the following.


W1

“The paper lacks controlled comparisons that incrementally introduce dynamic routing and selective correction under the same backbone, parameter count, and training settings.”

Response:

Thanks for the practical question. We designed and executed a strictly controlled ablation suite that holds the hierarchical backbone, trainable parameter count, prediction heads, primal-dual consistency supervision, benchmark refinement budget, and training schedule fixed, while incrementally introducing dynamic routing and selective correction.

To isolate Static versus Dynamic Routing under an identical parameter count, our Static Allocation variant feeds fixed training-set node embeddings ei\mathbf{e}_i into the exact same trainable router projection matrices W(l)\mathbf{W}^{(l)} (Eq. 19). Thus, Static Allocation and Dynamic Routing share the exact same trainable capacity, isolating the effect of conditioning the sparse assignment PlP_l on the current observation window XX.

Meanwhile, to isolate Random and Historical-Change Selection versus Learned Event Selection under an identical refinement budget (evaluated both at the 100% full benchmark budget in the table below and across constrained 10%–100% budgets in W2), we compare three active-mask selectors on the same dynamic routing backbone: (1) Random Selection (uniform random node selection and weighting per window), (2) Historical-Change Selection (nodes ranked and weighted by observed historical ramp magnitude), and (3) Learned Selection (our full event router).

The following Table presents a progressive ablation of our routing and selection mechanisms. Integrating window-dependent Dynamic Routing and Learned Selection in a step-wise manner consistently reduces MAE and RMSE across all four datasets. Specifically, when isolating the effect of routing (holding Learned Selection fixed), replacing Static Allocation with Dynamic Routing yields substantial MAE reductions: 3.56% on CA (16.01 → 15.44), 2.05% on GLA (16.58 → 16.24), 0.75% on SDWPF (0.2260 → 0.2243), and 1.78% on Electricity (0.2026 → 0.1990).

Controlled Incremental Ablation of Routing and Selection Mechanisms

VariantCAGLASDWPFElectricity
MAERMSEMAERMSEMAERMSEMAERMSE
Static Allocation + Random Selection16.1827.8116.7228.690.22680.39710.20410.3362
Static Allocation + Learned Selection (w/o Dyn. Grid)16.0127.5416.5828.460.22600.39580.20260.3338
Dynamic Routing + Random Selection15.7827.2116.4528.290.22560.39510.20150.3324
Dynamic Routing + Historical-Change Selection15.6026.9816.3928.180.22500.39410.20040.3308
AR-MGNet (Dynamic Routing + Learned Selection)15.4426.7616.2427.950.22430.39290.19900.3286
AuthorsOpenReview ↗

Author Responses (2/3)

W2

“Comparisons with random selection and selection based on the magnitude of historical changes under the same budget are missing.”

Response:

Thank you for this essential suggestion. To directly evaluate whether the learned router prioritizes the nodes most in need of correction, we added a Post-Prediction Oracle Selection-Quality Audit at the 50% refinement budget and a Multi-Budget Sweep from 10% to 100% on CA, GLA, SDWPF, and Electricity, comparing Learned Selection against Random Selection and Historical-Change Magnitude Selection.

1. Why Learned Selection Outperforms Historical-Change and Random Selection

As shown in the following Table, because AR-MGNet's learned event score si(l)s_i^{(l)} (Eq. 24) combines the cross-scale routing discrepancy ρi(l)\rho_i^{(l)} and bidirectional evidence discrepancy ΔEi(l)\Delta E_i^{(l)} with learned uncertainty ui(l)u_i^{(l)} and historical novelty νi(l)\nu_i^{(l)}, it achieves 58.4%–66.2% Top-Error Precision and captures 62.1%–71.4% of total inherited error while activating only 50% of the nodes, reducing 50%-budget forecast MAE on CA from 17.28 (Random) and 16.96 (Historical Change) down to 16.68.

Post-Prediction Oracle Selection-Quality Audit at 50% Refinement Budget

DatasetActive-Mask SelectorMAERMSETop-Error Oracle PrecisionInherited Error Captured
CALearned Event Router (Ours)16.68 ± 0.0627.82 ± 0.1264.8% ± 0.9%69.7% ± 1.1%
Historical-Change Magnitude16.96 ± 0.0728.15 ± 0.1357.9% ± 0.8%61.8% ± 0.9%
Random Selection17.28 ± 0.1028.52 ± 0.1650.0% ± 0.1%50.0% ± 0.1%
GLALearned Event Router (Ours)17.12 ± 0.0828.86 ± 0.1266.2% ± 0.8%71.4% ± 1.0%
Historical-Change Magnitude17.38 ± 0.0929.18 ± 0.1457.4% ± 0.9%61.2% ± 0.8%
Random Selection17.65 ± 0.1129.51 ± 0.1750.0% ± 0.1%50.0% ± 0.1%
SDWPFLearned Event Router (Ours)0.2258 ± 0.00120.3962 ± 0.001758.4% ± 1.0%62.1% ± 1.0%
Historical-Change Magnitude0.2269 ± 0.00140.3981 ± 0.001953.6% ± 1.4%56.4% ± 1.1%
Random Selection0.2279 ± 0.00210.3998 ± 0.002850.0% ± 0.1%50.0% ± 0.1%
ElectricityLearned Event Router (Ours)0.2142 ± 0.00110.3468 ± 0.001663.5% ± 0.9%68.2% ± 1.1%
Historical-Change Magnitude0.2176 ± 0.00120.3512 ± 0.001856.8% ± 1.0%60.4% ± 0.9%
Random Selection0.2215 ± 0.00170.3564 ± 0.002349.9% ± 0.2%49.9% ± 0.2%

2. Full Refinement Budget Sweep from 10% to 100%

We further sweep the refinement budget from 10% to 100% across 4 datasets. The Learned Router consistently achieves the highest oracle error capture and lowest forecast MAE. At η=0.50\eta = 0.50 and η=0.75\eta = 0.75, AR-MGNet captures 62.1%–71.4% and 84.6%–89.4% of total inherited error respectively, outperforming both Historical-Change and Random Selection. The results are shown as follows.

Refinement Budget Sweep: Forecast MAE and Oracle Error Capture

DatasetSelector Rule10% Budget MAE (Capture)25% Budget MAE (Capture)50% Budget MAE (Capture)75% Budget MAE (Capture)100% Budget MAE (Capture)
CALearned Event Router (Ours)17.65 (22.4%)17.22 (44.8%)16.68 (69.7%)16.14 (88.6%)15.44 (100.0%)
Historical-Change Magnitude17.92 (17.8%)17.51 (37.1%)16.96 (61.8%)16.36 (82.4%)15.60 (100.0%)
Random Selection18.21 (10.0%)17.84 (25.0%)17.28 (50.0%)16.62 (75.0%)15.78 (100.0%)
GLALearned Event Router (Ours)18.15 (23.1%)17.68 (46.2%)17.12 (71.4%)16.65 (89.4%)16.24 (100.0%)
Historical-Change Magnitude18.42 (17.2%)17.95 (35.7%)17.38 (61.2%)16.84 (84.8%)16.39 (100.0%)
Random Selection18.74 (10.0%)18.26 (25.0%)17.65 (50.0%)17.02 (75.0%)16.45 (100.0%)
SDWPFLearned Event Router (Ours)0.2278 (18.9%)0.2268 (39.5%)0.2258 (62.1%)0.2250 (84.6%)0.2243 (100.0%)
Historical-Change Magnitude0.2289 (15.4%)0.2279 (34.2%)0.2269 (56.4%)0.2258 (80.1%)0.2250 (100.0%)
Random Selection0.2298 (10.0%)0.2290 (25.0%)0.2279 (50.0%)0.2266 (75.0%)0.2256 (100.0%)
ElectricityLearned Event Router (Ours)0.2310 (21.5%)0.2234 (43.6%)0.2142 (68.2%)0.2058 (87.9%)0.1990 (100.0%)
Historical-Change Magnitude0.2348 (16.8%)0.2272 (36.4%)0.2176 (60.4%)0.2084 (82.5%)0.2004 (100.0%)
Random Selection0.2389 (10.0%)0.2315 (24.9%)0.2215 (49.9%)0.2112 (75.1%)0.2015 (100.0%)
AuthorsOpenReview ↗

Author Responses (3/3)

W3

“The metric definitions and test windows for the long-horizon results (Table 8) are unclear.”

Response:

We thank the reviewer for this insightful comment. The numerical difference between Table 1 and original Table 8 arises because the two tables evaluated different forecasting protocols and metric aggregation rules:

  • Metric Aggregation: Table 1 reports the 12-step Cumulative MAE (1:12), which pools the errors across all 12 future steps. In contrast, the original Table 8 reported the Endpoint MAE (@h) for multi-step autoregressive rollouts at horizons h∈{16,20,24,36,48}h \in \{16, 20, 24, 36, 48\}. These two metrics measure fundamentally different aspects of forecasting.
  • HieraMix: The jump from 16.45 in Table 1 to 33.14 in original Table 8 is because Table 1 reports the pooled 12-step Cumulative MAE, while original Table 8 reports the single-step Endpoint MAE at h=16h=16 during autoregressive rollout. Autoregressive feedback inherently amplifies boundary errors (steps 13–16), especially for models like HieraMix that mix multiscale features without an explicit macro-future solver to anchor the trajectory.
  • NEST: For NEST, its official multi-stage pre-warmup results are reported in original Table 8. Table 1 uses a unified from-scratch protocol for all models to avoid cross-protocol comparison.

To provide a fully transparent and reliable assessment of long-horizon advantages, we eliminated all cross-protocol differences by re-evaluating all models (AR-MGNet, HieraMix, PatchSTG, NEST) under a strictly unified protocol. All models and horizons use the same forecast origins, defined as those for which the entire 48-step future ground truth lies within the test split. The evaluation stride is one time step, and no forecast target crosses the test-split boundary. The following Table shows that AR-MGNet achieves the lowest Endpoint and Cumulative MAE at every horizon. This confirms that our long-horizon advantage is robust under this unified evaluation protocol.

Unified Autoregressive Endpoint/Cumulative MAE on GLA and CA

DatasetModelEndpoint / Cumulative MAE
MAE@16MAE@20MAE@24MAE@36MAE@48
GLAHieraMix33.14 / 20.1236.22 / 23.0838.78 / 25.5244.77 / 31.0849.33 / 35.19
PatchSTG25.63 / 19.8427.08 / 21.6227.92 / 23.1830.42 / 26.4232.43 / 28.76
NEST44.99 / 32.2048.71 / 35.2451.47 / 37.7856.67 / 43.3760.52 / 47.22
AR-MGNet20.61 / 17.2321.87 / 18.0822.91 / 18.8225.36 / 20.7126.55 / 22.07
CAHieraMix30.93 / 19.0433.97 / 21.8536.55 / 24.1842.65 / 29.5447.10 / 33.62
PatchSTG24.17 / 18.8625.52 / 20.5826.54 / 22.0528.10 / 24.6229.15 / 26.48
NEST42.95 / 30.6846.38 / 33.5649.12 / 36.0454.18 / 41.4257.86 / 45.15
AR-MGNet19.54 / 16.3420.72 / 17.1521.68 / 17.8623.95 / 19.6825.14 / 20.94

We hope these clarifications adequately address your concerns and demonstrate the soundness, contribution, and scientific rigor of our work. We are grateful for your feedback, which will help us improve the clarity of the final manuscript.

Reviewer_Evy7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a spatio-temporal forecasting model, which maps the observation nodes to functional regions (nodes) and do the predictions there, then guides the forecasting of observation nodes. The paper evaluate the method on multiple dataset and show its effectiveness.

Critical strengths
  1. The idea of coarse-to-fine make sense, especially doing the predictions on this functional region (nodes) and then pass the forecasting results back to fine-grained observation nodes to fix.
  2. The evaluations on dataset contain nodes number from hundreds to thousands shows that the cost is linear, which is a clear benefit of this approach.
Critical weaknesses
  1. The experiment results are inconsistent in the paper. For example, in Table 1 and Table 8, the performances of NEST method (most recent sota baseline) seem differ a lot even with same dataset. This makes the overall performance evaluation hard to interpret and understand. Also, it is unclear whether some comparing methods are re-trained (then what's the hyper-parameters for each) or just re-use their paper results.
  2. The model seems sensitive to the number of levels. for example, CA MAE is 15.44 with L = 3 but 23.44 with L = 4 in Table 4.
Questions for authors
  1. Could you explain the inconsistence between experiment results like discussed in the W1?
  2. More details about baseline experiment settings.
  3. All experiments seem to be done w/o mean and std, so only once? could you report such at least for some important ones?
6 comments
AuthorsOpenReview ↗

Author Responses (1/3)

We thank the reviewer for the thoughtful and constructive feedback. Below, we address your comments and questions, supplementing our analysis with additional experimental results.


W1 & Q1

“The experimental results between Table 1 and Table 8 are inconsistent”

Response:

Thanks for the question. The numerical difference between Table 1 and original Table 8 is due to different evaluation paradigms and metric aggregation rules:

  • Metric Aggregation: Table 1 reports the 12-step Cumulative MAE (1:12), which pools the errors across all 12 future steps. In contrast, the original Table 8 reported the Endpoint MAE (@h) for multi-step autoregressive rollouts at horizons h∈{16,20,24,36,48}h \in \{16, 20, 24, 36, 48\}. These two metrics measure fundamentally different aspects of forecasting.
  • Protocol Discrepancy: For NEST, its official multi-stage pre-warmup results are reported in original Table 8. Table 1 uses a unified from-scratch protocol for all models to avoid cross-protocol comparison. Our main empirical margin over the strongest identical-protocol baselines (HieraMix and PatchSTG) does not rely on NEST.

To eliminate any potential ambiguity caused by cross-protocol comparisons, we re-evaluated all competing methods (AR-MGNet, HieraMix, PatchSTG, and NEST) under a strictly unified protocol. All models and horizons use the same forecast origins, defined as those for which the entire 48-step future ground truth lies within the test split. The evaluation stride is one time step, and no forecast target crosses the test-split boundary. As shown in the following Table, we report both Endpoint MAE (@h) and Cumulative MAE (1:h). The results demonstrate that under an apples-to-apples comparison, AR-MGNet consistently achieves the lowest errors across all horizons.

Unified Autoregressive Endpoint/Cumulative MAE on GLA and CA

DatasetModelEndpoint / Cumulative MAE
MAE@16MAE@20MAE@24MAE@36MAE@48
GLAHieraMix33.14 / 20.1236.22 / 23.0838.78 / 25.5244.77 / 31.0849.33 / 35.19
PatchSTG25.63 / 19.8427.08 / 21.6227.92 / 23.1830.42 / 26.4232.43 / 28.76
NEST44.99 / 32.2048.71 / 35.2451.47 / 37.7856.67 / 43.3760.52 / 47.22
AR-MGNet20.61 / 17.2321.87 / 18.0822.91 / 18.8225.36 / 20.7126.55 / 22.07
CAHieraMix30.93 / 19.0433.97 / 21.8536.55 / 24.1842.65 / 29.5447.10 / 33.62
PatchSTG24.17 / 18.8625.52 / 20.5826.54 / 22.0528.10 / 24.6229.15 / 26.48
NEST42.95 / 30.6846.38 / 33.5649.12 / 36.0454.18 / 41.4257.86 / 45.15
AR-MGNet19.54 / 16.3420.72 / 17.1521.68 / 17.8623.95 / 19.6825.14 / 20.94

W1 & Q2

“More details about baseline experiment settings.”

Response:

Thank you for raising this important point. All baselines in Table 1 and Table 2 were re-trained from scratch on our datasets using their official repositories under identical window lengths, split boundaries, and normalization statistics. We have expanded Appendix B.2 in the revised manuscript to document the exact architectural parameters, optimizer configurations, and batch sizes for the listed baselines. The complete configurations are provided in Appendix B.2 of the revised paper.

AuthorsOpenReview ↗

Author Responses (2/3)

W2

“The model seems sensitive to the number of levels.”

Response:

We thank the reviewer for raising this critical question. We systematically diagnosed the L = 4 performance drop on CA in original Table 4 (15.44 at L = 3 vs. 23.44 at L = 4). Our analysis confirms that this sensitivity is caused by a structural spatial over-coarsening bottleneck under the standard partitioning schedule.

In the original Table 4 configuration on CA (N0=8,600N_0 = 8{,}600 sensors), the hierarchy used an standard coarsening schedule (8 → 4 → 4), which compressed the spatial graph into 8,600 → 1,075 → 269 macro-cells at L = 3, and further down to only N3=68N_3 = 68 macro-cells at L = 4. While 269 functional macro-regions preserve distinct freeway corridors and directional bottlenecks across California, compressing 8,600 statewide road sensors into just 68 macro-cells forces each macro-cell to average over more than 126 heterogeneous sensors spanning opposing traffic directions and separate metropolitan areas. This excessive spatial compression over-smooths regional traffic dynamics in the prolonged macro forecast Yˉ(l)\bar{\mathbf{Y}}^{(l)}, creating a severe spatial information bottleneck. By contrast, on KnowAir (original Table 4), where atmospheric pollutant transport is spatially smoother and less topologically fragmented, moving from L = 3 to L = 4 preserves accuracy (14.84 vs. 14.71 MAE; 22.89 vs. 23.00 RMSE).

To verify that the L = 4 degradation stems from spatial over-coarsening rather than optimization failure or router collapse, we modified the spatial partitioning scheme on CA to use a balanced 4-fold spatial coarsening schedule (8,600 → 2,150 → 538 → 135 macro-cells) so that the coarsest level at L = 4 retains 135 functional macro-cells instead of 68, which improves CA test performance from 23.44 / 37.30 to 17.82 MAE / 28.19 RMSE (reducing the gap relative to L = 3 from +8.00 MAE down to just +2.38 MAE).

Effect of the Number of Grid Levels of AR-MGNet on CA

CA settingGrid sizesMAERMSE
L = 1860035.2555.28
L = 28600 → 107522.9835.86
L = 38600 → 1075 → 26915.4426.76
L = 4 (Original Table 4)8600 → 1075 → 269 → 6823.4437.30
L = 4 (Another Option)8600 → 2150 → 538 → 13517.8228.19
AuthorsOpenReview ↗

Author Responses (3/3)

Q3

“All experiments seem to be done w/o mean and std.”

Response:

We thank the reviewer for this insightful comment. We have re-run our core experiments across 3 independent random seeds (Seeds 42, 43, 44) with strong baselines and report the Mean ± Sample Standard Deviation. Across SD, GBA, GLA, and CA, AR-MGNet exhibits tight training repeatability (MAE sample standard deviation between 0.05 and 0.08, i.e., < 0.5% of the mean). Across the five non-traffic datasets, AR-MGNet's sample standard deviations are 0.0010–0.0012 on SDWPF, Electricity, and Solar-Energy, 0.06 on KnowAir, and 0.03 on UrbanEV, confirming that its cross-domain improvements (up to 19.0% over the strongest baseline) are highly stable across initializations. The below Tables and Appendix E.1 give the complete LargeST and cross-domain results below. All entries are mean ± sample standard deviation over seeds 42, 43, and 44.

Three-Seed Evaluation on LargeST

ModelSDGBA
MAERMSEMAPE (%)MAERMSEMAPE (%)
HieraMix14.80 ± 0.0925.06 ± 0.159.22 ± 0.0717.73 ± 0.0730.67 ± 0.1612.71 ± 0.11
PatchSTG15.93 ± 0.1128.19 ± 0.189.69 ± 0.0918.94 ± 0.1233.02 ± 0.2013.17 ± 0.14
NEST24.41 ± 0.1439.09 ± 0.2117.11 ± 0.1523.26 ± 0.1637.04 ± 0.1421.49 ± 0.22
AR-MGNet14.64 ± 0.0724.94 ± 0.128.94 ± 0.0517.62 ± 0.0629.87 ± 0.1310.95 ± 0.07
ModelGLACA
MAERMSEMAPE (%)MAERMSEMAPE (%)
HieraMix16.45 ± 0.1128.03 ± 0.149.53 ± 0.0615.55 ± 0.0827.05 ± 0.1310.55 ± 0.07
PatchSTG17.53 ± 0.1329.72 ± 0.1910.02 ± 0.0816.69 ± 0.1028.58 ± 0.1711.50 ± 0.09
NEST28.38 ± 0.1345.38 ± 0.1718.11 ± 0.1627.16 ± 0.4443.87 ± 0.5721.19 ± 0.31
AR-MGNet16.24 ± 0.0827.95 ± 0.119.32 ± 0.0415.44 ± 0.0526.76 ± 0.1110.40 ± 0.05

Three-Seed Evaluation across Five Non-Traffic Datasets

ModelKnowAirUrbanEVElectricity
MAERMSEMAERMSEMAERMSE
STID16.66 ± 0.0525.91 ± 0.044.89 ± 0.0311.28 ± 0.030.2355 ± 0.00010.3819 ± 0.0004
HieraMix15.90 ± 0.1524.38 ± 0.025.19 ± 0.0211.24 ± 0.010.2437 ± 0.00070.3909 ± 0.0007
NEST16.02 ± 0.0724.43 ± 0.055.35 ± 0.0411.25 ± 0.010.2350 ± 0.00030.3755 ± 0.0016
AR-MGNet14.84 ± 0.0622.89 ± 0.094.36 ± 0.0310.80 ± 0.040.1990 ± 0.00100.3286 ± 0.0015
ModelSDWPFSolar-Energy
MAERMSEMAERMSE
STID0.2348 ± 0.00240.4092 ± 0.00300.2182 ± 0.00400.4429 ± 0.0137
HieraMix0.2335 ± 0.00240.3995 ± 0.00480.2322 ± 0.01120.4554 ± 0.0253
NEST0.2342 ± 0.00040.3992 ± 0.00440.2110 ± 0.00470.4483 ± 0.0035
AR-MGNet0.2243 ± 0.00110.3929 ± 0.00160.1709 ± 0.00120.3634 ± 0.0018

We sincerely thank you for your encouraging evaluation and thoughtful comments. Your recognition of our work’s contributions is highly motivating, and your suggestions have helped us further strengthen the clarity and rigor of the manuscript.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Reviewers,

We sincerely thank all the reviewers (Evy7, xnTy, Eisp) for their valuable feedback. We are glad that the reviewers appreciated the effectiveness and consistency of our proposed framework (Evy7, xnTy), the extensiveness of our experiments (Evy7, xnTy, Eisp), the solidness of experimental results (Evy7, xnTy, Eisp), and the overall quality of our paper's writing (Eisp).

We have made every effort to faithfully address your comments in the responses. As suggested by the reviewers, we add:

  • Unified re-evaluation of long-horizon forecasting protocols: Added in Table 12 in Appendix E.3 to eliminate cross-protocol ambiguity, as requested by Reviewers Evy7 and xnTy.
  • Expanded baseline experiment settings: Documented in Appendix B.2, as requested by Reviewer Evy7.
  • Structural analysis of grid-level sensitivity and balanced L=4 partition on CA: Detailed in Table 6 in Section 5.2, as requested by Reviewers Evy7 and Eisp.
  • 3-seed stability assessment and bootstrap confidence intervals: Results are shown in Tables 9 and 10 in Appendix E.1, requested by Reviewers Evy7 and Eisp.
  • Improvement claims against baselines: Discussed in Appendix D, as requested by Reviewer Eisp.
  • Controlled incremental ablation of routing and selection mechanisms: Included in Table 3 in Section 5.2 under identical backbones, parameter counts, and supervision, as requested by Reviewers xnTy and Eisp.
  • Post-prediction oracle selection-quality audit and refinement-budget sweep: Detailed in Table 4 in Section 5.2, across 50% and 10%–100% budgets on four datasets, requested by Reviewers xnTy and Eisp.

We have incorporated the suggested modifications in the revised version, which are highlighted in blue. If you are satisfied with the revisions, we kindly request your approval to consider an improved score.

Thanks for all the reviewers' time again.

Best regards,

Authors

AuthorsOpenReview ↗

General Summary by Authors

Dear AC and Reviewers,

 As the discussion phase comes to a close, we would like to provide a brief general summary of our paper and express our sincere gratitude for your time and constructive feedback. 

Core Contributions and Highlights

Our work proposes AR-MGNet, which reformulates spatio-temporal forecasting as a bidirectional macro-micro reconciliation process. The key highlights are:

  1. A Novel Paradigm: We explicitly separate efficient coarse-scale macro-future solving from selective fine-scale innovation recovery, ensuring both global coherence and local precision.
  2. Dynamic Multi-Grid Learning: We dynamically partition functional grids for each input window, constructing sparse, relation-aware adjacencies that adapt to current spatial regimes.
  3. Consistency-Constrained Optimization: We jointly supervise hierarchical forecast fields to separate active innovations from inherited forecasts.
  4. Efficiency & Scalability: The sparse routing and contraction core scales linearly with the number of nodes, demonstrating high practical value.  Extensive Experimental Validation

We conducted comprehensive evaluations across 9 diverse benchmarks (4 LargeST traffic regions, plus 5 non-traffic domains: wind, air quality, EV charging, electricity, and solar). AR-MGNet achieved the best performance across traffic and cross-domain datasets. 

During the revision process, we further significantly strengthened the paper's rigor by adding:

  • Strictly controlled incremental ablations isolating routing and selection mechanisms.
  • Post-prediction oracle selection-quality audits across various refinement budgets.
  • 3-seed statistical evaluations with mean ± standard deviation and bootstrap confidence intervals.  We sincerely thank the Area Chair for managing this process and for your initiative to maintain high review quality. We also deeply appreciate all the reviewers for their rigorous reviews and valuable suggestions, which have greatly helped us solidify the evidence of our claims and significantly improve the manuscript.  We have done our absolute best to address all concerns while keeping the core methodology and primary conclusions intact. Thank you again for your time and guidance. 

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.