- 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
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
| Domain | AR-MGNet (Ours) | HieraMix | Best baseline | |||
|---|---|---|---|---|---|---|
| MAE | RMSE | MAE | RMSE | MAE | RMSE | |
| Solar-Energy | 0.1709 | 0.3634 | 0.2322 | 0.4554 | 0.2110 (NEST) | 0.4347 (TimeMixer) |
| Electricity | 0.1990 | 0.3286 | 0.2437 | 0.3909 | 0.2295 (PatchSTG) | 0.3699 (PatchSTG) |
| UrbanEV | 4.36 | 10.80 | 5.19 | 11.24 | 4.89 (STID) | 11.24 (HieraMix) |
| KnowAir | 14.84 | 22.89 | 15.90 | 24.38 | 15.90 (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).
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 ( 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 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 .
-
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 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 setting | Grid sizes | MAE | RMSE |
|---|---|---|---|
| L = 1 | 8600 | 35.25 | 55.28 |
| L = 2 | 8600 → 1075 | 22.98 | 35.86 |
| L = 3 | 8600 → 1075 → 269 | 15.44 | 26.76 |
| L = 4 (Original Table 4) | 8600 → 1075 → 269 → 68 | 23.44 | 37.30 |
| L = 4 (Another Option) | 8600 → 2150 → 538 → 135 | 17.82 | 28.19 |
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.
Author Responses (1/3)
Thank you very much for your insightful and constructive feedback on our submission. The following are our detailed responses.
W1
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
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