acceptodds
Under review as a conference paper at ICLR 2027

TOPODUET: ONE TOPOLOGY, TWO RELATIONAL FIELDS FOR SPATIO-TEMPORAL FORECASTING

Abstract

Spatio-temporal graph forecasters typically encode observations and temporal variations in a single latent tensor, although locations that look alike need not change alike. We introduce TOPODUET, which separates these relational roles into a Configuration Field \(L\) induced by absolute observations and a Transition Field \(I\) induced by signed first- and second-order variations. The Dual-Field Spatio-Temporal Encoder learns independent temporal and spatial relations for the two fields. Field-Adaptive Road Attention supplies both fields with forward–backward road support while learning field- and layer-specific prior strengths, and Progressive Cross-Field Coupling exchanges their evidence after each block without collapsing their identities. A horizon-adaptive dual-path readout then balances nonlinear correction with historical persistence. Our structural analysis quantifies the distortion caused by forcing one shared relation, identifies the minimal direction-complete closure of a directed road prior, and characterizes when a common soft prior yields identical field relations. Experiments on eight traffic forecasting datasets show that TOPODUET ranks first on all 12 MAE/RMSE comparisons across six graph-based highway datasets, reducing the best competing error by up to 4.9% in MAE and 7.3% in RMSE. The same architecture also transfers effectively across the T-Drive and CHIBike grid benchmarks, extending the dual-field formulation from sensor graphs to urban-flow fields. Controlled ablations further show that removing field separation increases MAE/RMSE by 0.10/0.62 on PeMS08 and 0.28/0.95 on T-Drive, supporting the value of modeling configuration and transition relations separately.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatio-Temporal Forecasting, Spatio-Temporal Graph Neural Networks, Graph Attention, Dual-Field Representation, Road Network Modeling, Time Series Forecasting

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_v1V6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TopoDUET, which employs a "Dual-Field Spatio-Temporal Encoder" to model two distinct relational fields, and a "Horizon-Adaptive Dual-Path Readout" to integrate non-linear residual predictions with linear historical extrapolations. The authors conduct experiments across multiple datasets, reporting strong results on most metrics, and provide ablation studies alongside an analysis of the relational mechanisms. While the overall design logic is coherent, the current evidence is insufficient to definitively attribute the performance gains primarily to the "relational decoupling" mechanism emphasized by the authors.

Critical strengths
  1. The experiments have a broad scope.

  2. The structure of the methodology aligns clearly with the three challenges raised in the paper.

Critical weaknesses
  1. The authors describe the "single latent carrier" as a "fundamental challenge," yet the main text provides no direct evidence that existing models actually suffer from significant "relational aliasing." The current arguments rely primarily on conceptual examples rather than empirical analysis of conflicts within the relational matrices of actual models.
  2. The "Transition Field" is defined simply as a combination of first-order and second-order differences; however, there is no justification for why this specific "first-order + second-order" formulation constitutes the most appropriate representation for transitions.
  3. The model strictly follows a "temporal → spatial" processing order, but no evidence is provided to demonstrate that this sequence is superior to a "spatial → temporal" approach.
  4. The paper lacks comparisons involving kk-hop (1/2/3/K-hop) neighborhoods or shortest-path decay mechanisms.
  5. Information for reproducibility is severely lacking; simply specifying the GPU, optimizer, and training strategy is insufficient to fully replicate the results.
  6. There is a lack of sufficient evidence regarding statistical significance; for the vast majority of baselines, only a single mean value is reported.
  7. Ablation studies are conducted exclusively on the PeMS08 and T-Drive Inflow datasets—a scope insufficient to establish the generalizability of the ablation findings.
  8. I will consider adjusting my score based on the authors' rebuttal.
Questions for authors

Q1. How do you disentangle the gains resulting from increased model capacity?

Q2. Why must the "first-order + second-order difference" be used for the Transition Field?

Q3. Are the marginal improvements shown in Tables 2–3 statistically significant?

Q4. Please explain and verify the concept of "reverse support" within FARA.

Reviewer_ePr3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

TOPODUET is a spatio-temporal forecasting architecture that processes absolute observations and signed first- and second-order temporal differences through separate configuration and transition fields. Each field has independent temporal and spatial attention, while a shared road-derived prior, field-specific prior strengths, and gated cross-field updates coordinate their representations. A horizon-adaptive readout combines nonlinear residual predictions with a linear history-based path. The paper evaluates the model on six highway datasets and two urban-grid datasets, reports first-place results on all 12 highway MAE/RMSE comparisons, and presents ablations, attention analyses, and structural propositions to support the design.

Critical strengths
  1. The architectural motivation is coherent. Similar observation levels need not imply similar future changes, so separately processing levels and temporal differences provides a reasonable inductive bias. Independent pathways with progressive interaction offer a concrete implementation of this distinction and a meaningful hypothesis to evaluate.

  2. The experimental scope spans different forecasting settings. The datasets include traffic flow and speed, sensor graphs and urban grids, and substantially different network sizes. This breadth is useful for assessing the applicability of the proposed representation across signals and spatial supports.

Critical weaknesses
  1. Research integrity: baseline results in Tables 2 and 3 appear to have been altered.

I compared every baseline entry in Tables 2 and 3 that has a counterpart in a published paper, 212 entries in total. The sources are PDFormer (AAAI'23, Tables 2–3), STG-NCDE (AAAI'22, Tables 3–4), STAEformer (CIKM'23, Table 3), RAST (AAAI'26, arXiv:2508.16623v2, Table 1), and MFormer (Biomimetics'26, Tables 4–6). Only 53 entries match the published values. The remaining 159 differ in systematic patterns. The paper does not identify whether baseline numbers are cited or rerun, and the supplementary material explicitly excludes baseline code.

Every traceable published result in Table 2 that outperforms TOPODUET is printed as worse than TOPODUET. There are exactly six such entries:

Dataset and metricTOPODUETBaselinePublished valuePrinted in Table 2
PeMS04 RMSE30.41PDFormer29.96531.24
PeMS04 RMSE30.41STAEformer30.1830.97
PeMS04 RMSE30.41RAST29.9330.55
PeMS07 RMSE32.96STAEformer32.6033.18
PeMS07 RMSE32.96PDFormer32.87033.94
PeMS07 RMSE32.96RAST32.7332.97

The RAST row is particularly concerning. Its other six entries match the camera-ready version exactly. Only the two cells where RAST outperforms TOPODUET differ, with PeMS07 RMSE printed just 0.01 above TOPODUET. Neither arXiv version reports 30.55 or 32.97. Using the published values, the abstract's claim that TOPODUET “ranks first on all 12 MAE/RMSE comparisons” does not hold.

Older baseline results exhibit near-constant rescaling. In Table 3, the 40 entries for DCRNN, STGCN, STGODE, STSGCN, and TFormer have a mean ratio of 0.9795 to PDFormer's published values, with a standard deviation of 0.0011. All 20 T-Drive ratios fall between 0.9786 and 0.9804. In Table 2, all 72 PeMS03/04/07/08 entries for nine older baselines are 1.0–3.4% below the STG-NCDE paper's values, while the corresponding PeMSD7(M)/(L) entries from the same source are copied verbatim. These patterns require an explicit account of how the numbers were produced. The Table 3 rescaling strengthens those older baselines; its relevance is the provenance of the printed results.

Some rows appear misidentified. Table 3's “ASTGCN (Guo et al. 2019)” tracks ASTGNN (Guo et al. 2021), a different model, multiplied by approximately 0.98. The exception is CHIBike-Inflow RMSE, which is printed as 5.09 rather than approximately 5.70 and highlighted as the best result. Table 2's “TFormer (Cai et al. 2020)” instead tracks TFormer by Yan and Ma, as reported in PDFormer, multiplied by approximately 1.03.

These issues directly compromise the main empirical comparison and are decisive to my recommendation. I request AC and ethics-reviewer examination of the baseline provenance.

  1. The contribution of field separation is not adequately isolated, and implementation attribution is incomplete.

The ablations do not establish whether the gains arise specifically from separating configuration and transition relations, rather than additional capacity, difference features, or a generic two-stream architecture. The decisive controls are a single-stream model receiving [x, Δx, Δ²x], a capacity-matched single-stream model, and a two-stream model receiving identical inputs. Table 4 also omits variability and reports Full-model results different from the main-table means without explaining the run-selection convention.

The structural propositions do not establish forecasting superiority: attention-matrix distortion does not imply increased prediction error, and nonzero attention disagreement does not establish the intended semantic specialization.

Implementation provenance also needs correction. The supplementary data loader is a near-verbatim copy of STAEformer's lib/data_prepare.py, and the README states that the model was developed inside the STAEformer repository. The paper does not acknowledge this reuse or clearly distinguish the inherited components from the proposed implementation.

Questions for authors
  1. Clarificationof baseline provenance: For every baseline entry in Tables 2 and 3, please identify its source paper and table, or state that it was rerun and provide the corresponding code, configuration, and logs. What produced the systematic rescalings and the ASTGCN/ASTGNN and TFormer row identities?

  2. Clarification of RAST discrepancies: Why are PeMS04 and PeMS07 RMSE reported as 30.55 and 32.97 when neither RAST arXiv version contains these values, while the other six entries match its camera-ready table? What records support these two numbers?

  3. Clarification of implementation and ablation provenance: Which components were inherited from STAEformer, and where is that reuse acknowledged? Which seeds and checkpoints produced Table 4's Full results, and what are the exact implementations, parameter counts, and per-seed results of the field-removal and field-separation variants?

  4. Evidence supporting the current central claim: What capacity- and training-matched evidence establishes that configuration–transition separation itself improves forecasting over the three controls described above?

Reviewer_izBrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TOPODUET, a dual-field framework for spatio-temporal forecasting. It separates absolute observations and signed first- and second-order temporal differences into configuration and transition fields, respectively, to address what the authors term relational aliasing. The Dual-Field Spatio-Temporal Encoder (DFSTE) applies separate temporal and spatial attention pathways to the two fields. Field-Adaptive Road Attention (FARA) introduces a shared forward–backward road affinity as an additive spatial attention bias, with independently learned strengths for each field and encoder layer. The prior’s support forms the minimal direction-complete extension of the forward support. Progressive Cross-Field Coupling (PCFC) enables gated residual exchange after each encoder block, while a horizon-adaptive dual-path readout combines nonlinear residual prediction with a linear projection of historical observations. Experiments cover 8 datasets. Among the compared baselines, TOPODUET achieves the lowest reported mean MAE and RMSE on all 6 highway datasets, leads on 3 of the 4 T-Drive metrics, and trails several stronger baselines on CHIBike. Ablations and attention analyses provide supporting evidence for the design.

Critical strengths
  1. TOPODUET performs well across six highway benchmarks covering both traffic flow and traffic speed, with networks ranging from 170 to 1,026 nodes. It achieves the lowest reported mean MAE and RMSE among the compared baselines on every dataset. The reported standard deviations across three seeds are small. The breadth of these results supports the architecture’s usefulness across different signal types and graph sizes, although some gains over the strongest baselines are modest.
  2. The distinction between relations induced by observation levels and those induced by temporal changes gives the architecture a clear rationale. Separate input projections, field-specific embeddings, and dedicated spatio-temporal encoding pathways make this distinction explicit, while gated interaction allows the fields to exchange information after each block. Each field models both temporal and spatial dependencies, so the separation follows the source of relational evidence. This connection between the motivation and the implementation makes the design easy to interpret.
Critical weaknesses
  1. The reliability of some performance gains needs clearer experimental documentation. The paper reports results across a broad set of benchmarks, but does not clarify whether the strongest baselines were rerun or taken from prior publications, or whether they used the same data splits and evaluation protocol. Table 2 also reports multi-seed statistics only for TOPODUET, while some margins are small. These omissions do not negate the overall results, but reduce my confidence in the gains over the strongest baselines.
  2. The ablations generally support the design, but the contributions of individual components remain kind of unclear. The implementation of “w/o Field Separation” is not described in enough detail to assess whether input information and model capacity are adequately controlled. Removing PCFC also improves RMSE on PeMS08, although MAE worsens, and the ablation table does not report variability across runs. This makes it difficult to distinguish consistent component benefits from metric trade-offs or random variation. The concern primarily limits confidence in the proposed mechanism, rather than the overall usefulness of the dual-field framework.
Questions for authors
  1. Were the strongest baseline results obtained from reruns or prior publications? Did they use the same data splits, forecast horizons, and target masks? For the closest competitors, is variability across runs available to help assess the reliability of the smaller gains?
  2. Does “w/o Field Separation” retain the same observation and difference inputs, with a parameter count comparable to the full model? Also, is the improvement in RMSE alongside worse MAE after removing PCFC on PeMS08 consistent across seeds, and how do the authors interpret this trade-off?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.