acceptodds
Under review as a conference paper at ICLR 2027

How Much Structural Abstraction Does Traffic Forecasting Need? A Benchmark for Controllable Multimodal Representations

Abstract

Traffic forecasting combines temporal, visual, and graph representations, yet it remains unclear how much structural abstraction should be provided to a model rather than inferred from raw observations. Benchmarks make this question difficult to isolate because representation type, graph construction, and model architecture are usually varied together. We introduce GB-Traffic, a representation-centric benchmark that aligns temporal signals, multiscale visual encodings, and functional temporal graphs while controlling structural abstraction from coarse to fine. All representations are constructed from historical observations, allowing abstraction level to be studied without changing the prediction target. The benchmark covers six datasets spanning highway, urban-road, bus rapid transit, and metro systems under a unified forecasting protocol and evaluates granularity controllability, functional-edge verification, multimodal complementarity, structural utility across backbones, cross-granularity transfer, and zero-shot cross-domain transfer. Results show that increasing structural resolution is not uniformly beneficial. Intermediate abstraction provides a consistent balance between local detail and structural coherence, while functional temporal graphs provide complementary information beyond physical connectivity across settings. Cross-domain results show that transferability depends on the representation regime. These observations establish structural abstraction as an explicit experimental variable in spatio-temporal forecasting rather than a fixed preprocessing decision. Code, representations, and evaluation resources are available at the anonymous repository: https://anonymous.4open.science/r/GB-Traffic-5FCB.

Primary area:
datasets and benchmarks
Keywords:
Structural Abstraction, Multimodal Representation Learning, Functional Temporal Graphs, Spatio-Temporal Forecasting, Traffic 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_GmY8OpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces GB-Traffic, a representation-centric benchmark for studying how different levels of structural abstraction affect traffic forecasting. Instead of fixing the graph representation and only comparing forecasting models, the paper aligns raw temporal signals, multiscale visual representations, and coarse-to-fine T-GRIG functional graphs under the same historical window and prediction target. The benchmark covers six traffic datasets and evaluates granularity control, functional-edge quality, multimodal complementarity, structural utility across several forecasting backbones, cross-granularity transfer, and zero-shot cross-domain transfer. The main observation is that finer structural resolution is not always better, and an intermediate abstraction level often provides a stronger balance between local detail and structural coherence.

Critical strengths

The strongest contribution is the decision to make structural abstraction itself a controllable benchmark variable. This is a more interesting contribution than simply introducing another forecasting architecture, because it enables questions about coarse, intermediate, and fine structural representations under a fixed prediction task. The paper also makes a serious effort to separate different types of evidence: functional-edge verification is not treated as sufficient proof of forecasting utility, multimodal gains are separated from structural-source replacement, and the zero-shot results are explicitly presented as a stress test rather than as universal domain generalization. This relatively careful claim framing makes the benchmark more convincing.

Critical weaknesses

My main concern is that the paper asks a broad question about how much structural abstraction traffic forecasting needs, but the strongest predictive evidence is still tied to one particular abstraction pipeline. Coarse, medium, and fine representations are all generated by T-GRIG, and the downstream comparisons depend on the same region-to-sensor interface. As a result, the paper demonstrates that an intermediate abstraction level works well within this specific representation design, but it is less clear whether the conclusion extends to structural abstraction more generally. In addition, the experimental coverage is somewhat uneven: the multimodal forecasting results are mainly reported on PeMS04 and METR-LA, while some of the broader six-dataset evidence comes from graph statistics and transfer diagnostics. The authors acknowledge these scope boundaries, but I think the main text could be more conservative in presenting the generality of the conclusion.

Questions for authors

The main question I would like the authors to address is how much of the observed granularity effect is attributable to T-GRIG itself and how much may come from the region-to-sensor adapter. Since the paper explicitly notes that the structural-interface results measure the joint graph-plus-adapter pipeline, it would be helpful to clarify whether the same coarse-versus-medium-versus-fine trend remains under a simpler or fixed aggregation rule. I would also like the authors to clarify how broadly they intend the “intermediate abstraction is preferable” conclusion to be interpreted: is this mainly an empirical property of T-GRIG under the current benchmark, or do they view it as a more general statement about traffic forecasting representations?

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

This paper introduces GB-Traffic, which aims to systematically investigate the extent to which structural abstraction is required in models for traffic forecasting. The benchmark evaluates performance across six key questions. Experiments indicate that a medium level of granularity generally achieves an optimal balance between preserving local details and maintaining structural consistency. Overall, however, the benchmark remains heavily reliant on the authors' proprietary T-GRIG representation framework; consequently, the conclusions regarding the "degree of structural abstraction" lack independent validation across alternative architectural approaches.

Critical strengths
  1. Attempt to decouple the representation backbone from the forecasting backbone.
  2. Use a 12-step historical window and a 12-step forecasting target consistently across all six datasets.
Critical weaknesses
  1. As a benchmark, it lacks sufficient diversity in its structural abstraction methods.
  2. The term "multimodal" is somewhat of an overstatement; these are more accurately described as "multi-view" or "multi-representation" approaches, rather than "multimodality" in the traditional, independent sense.
  3. The comparison between the physical graph and T-GRIG does not strictly isolate "structure" as the only variable being changed.
  4. All tasks are essentially fixed at a 12-step input to 12-step output configuration (covering a maximum prediction horizon of 60 minutes). Consequently, the scope implied by the title—"How Much Structural Abstraction Does Traffic Forecasting Need?"—is broader than what the experiments actually cover.
  5. In T-GRIG, nodes correspond to regions within a transformed temporal representation rather than physical sensors, yet the prediction targets remain sensor-level values.
  6. The labels "coarse," "medium," and "fine" may not represent the same degree of abstraction across different datasets.
  7. The study should incorporate more advanced spatio-temporal Transformers, robust temporal-only models, and strong non-graph-based baselines to address the question of whether such "structural abstraction" is still necessary for state-of-the-art models.
  8. Many of the conclusions in this benchmark paper are based on performance margins of only a few percentage points—or even less. Relying solely on single point estimates is insufficient to determine whether findings (such as those regarding "medium granularity") are stable.
  9. There is a lack of analysis concerning the efficiency and scalability of the representation construction process.
  10. I will determine my final score based on the authors' rebuttal.
Questions for authors

Q1. How can it be demonstrated that the "medium is best" finding is not a phenomenon unique to T-GRIG?

Q2. How does the comparison between the physical graph and T-GRIG account for (or isolate) the influence of sample-dependent information?

Q3. Does the functional-edge validation mentioned in Q2 suffer from a circular reasoning (self-justification) issue?

Q4. Please report the statistical significance of the main results based on multiple random seeds.

Q5. What is the computational cost of constructing T-GRIG for real-world, large-scale transportation networks?

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

The paper introduces GB-Traffic, a benchmark built from six traffic datasets (PeMS04, PeMS08, METR-LA, XMBRT, HZMetro, BJMetro). The task is to predict the next 12 five-minute steps from the last 12. For each sensor and each one-hour window, the benchmark provides the raw sequence, an image built from this sequence, and three graphs built by T-GRIG. This method splits the image into regions of similar pixels, and each region becomes a node. A threshold on this similarity controls the graph size. The paper studies six questions, for example, which graph size works best and whether adding T-GRIG to the time series and the image improves forecasting. It reports that the medium graph works best and that adding T-GRIG lowers the forecasting error in every comparison of Tables 5-7, typically by several percent.

Critical strengths

The question is relevant: which representation of the same input window helps forecasting, i.e., the raw sequence, an image, or a graph, and how detailed this graph should be. Providing all these representations for the same windows, with the same split and target, is a reasonable design for such a study.

Critical weaknesses

1. The method and the experiments are not described concretely.

  • The paper presents T-GRIG graphs as a representation of traffic structure. However, each graph is built from the image of one sensor's last hour, which gives one separate graph per sensor and time window. Its nodes are regions of this image, so the graph describes one sensor's hour and contains no information about the road network. Forecasting models such as STGCN and Graph WaveNet need one input per sensor, and the paper obtains it with a "region-to-sensor adapter". Appendix C.2 introduces this adapter only as a function symbol and does not say what it computes. All results in Section 4.4 depend on it, but the paper neither describes it nor reports results without it (lines 933-935).
  • The paper does not say which model produces the rows of Table 5 or how zero-shot transfer works between datasets with different numbers of sensors (Section 4.6). It also gives no training settings, which it treats as "configuration-level records" (lines 512-513).
  • Some statistics do not fit the definition of T-GRIG. Lines 316-318 state that "among long-distance functional edges above 5 km on PeMS04, 82% align morning peaks within 10 minutes". However, a T-GRIG edge connects two regions of an image built from one sensor's last hour. These regions have no location on the road, so the distance between them cannot be measured in kilometers. Moreover, finding a morning peak requires data for the whole morning, while each input covers one hour. The paper does not explain how these statistics were computed. For the same reason, it is unclear whether the "pairs" in Table 4 are pairs of image regions or pairs of sensors.
  • Many sentences only limit the paper's own claims, for example "should not be interpreted as..." and "rather than a...", while the method itself stays unexplained. Appendices K-P contain tables about file formats and versioning rules without any results.

2. The experiments do not support the main claims of the paper.

The paper claims that the medium graph works best and that adding T-GRIG improves forecasting. The experiments do not support these claims for the following reasons:

  • No result comes with a standard deviation (Appendix O), while many gains are only a few percent. For example, the medium graph beats the fine one by 19.47 vs 19.78 MAE (Table 3).
  • The forecasting comparisons behind these claims use one or two datasets. The comparison of thresholds (Table 3), the comparison of forecasting models (Table 6) and the ablation of T-GRIG components (Table 8) use only PeMS04, and Tables 5 and 7 use PeMS04 and METR-LA. The other four datasets of the benchmark (PeMS08, XMBRT, HZMetro, BJMetro) appear in none of these comparisons.
  • Some numbers repeat across experiments that, according to lines 270-272, use different inputs. On PeMS04, the 60-minute MAE is 19.47 for the model with time series, image and T-GRIG (Table 5), for Graph WaveNet with the "T-GRIG-derived" graph (Tables 6 and 7), and for the medium graph in the comparison of thresholds (Table 3). Similarly, the METR-LA MAE of 4.95 appears in both Table 5 and Table 7.
  • On METR-LA at 60 minutes, Table 7 reports an MAE of 5.10 for Graph WaveNet with its adaptive graph, while the original Graph WaveNet paper reports 3.53 for the same input length and horizon. The setups differ in the split (60:20:20 vs 70:10:20), and the paper does not say whether zeros, which mark missing values in METR-LA, count among the "valid targets" of MAE (lines 267-268). Therefore, it is unclear whether the gains of T-GRIG would hold against baselines evaluated as in the original paper.

3. The released code does not match the paper.

The anonymous repository linked in the abstract contains two code files. Model/GAF_op.py builds the images and trains a ResNet-18 on the images alone, and Model/GB.py builds the T-GRIG graphs. The benchmark data, the forecasting models and the "region-to-sensor adapter" (Appendix C.2) are not released, so none of the forecasting results (Tables 3 and 5-8, Figures 3-4) can be reproduced. The mentioned ResNet-18 on images alone has no matching row in Table 5. Moreover, the released code differs from the paper:

  • The images built by Model/GAF_op.py have no MTF channels, and their time scales are 5, 15, 30 and 60 minutes instead of 5, 10, 30 and 60 (Model/GAF_op.py:88-109).
  • The T-GRIG code in Model/GB.py has no "temporal-axis weighting" (the parameter α\alpha is stored but never used), although Table 8 reports an ablation of it. It also merges regions in random order (Model/GB.py:51) instead of the gradient-based order of Algorithm 1, and its node features are per-channel means and standard deviations (Model/GB.py:65-73) instead of the 19 features described in the paper (lines 253-255).
  • The training script in Model/GAF_op.py computes the normalization statistics on all data, including the test period (Model/GAF_op.py:197), contrary to lines 171-178. Its errors are multiplied by the standard deviation of the data, because it rescales targets that were never normalized (Model/GAF_op.py:173-174,200). Finally, it takes the input and target windows of PeMS04 and PeMS08 from the wrong axis of the data array (Model/GAF_op.py:68-69): on PeMS04, the "history" is sensors 0-11 and the "future" is sensors 12-23 at the same time step.
  • Figure 2(d-f) seems to be produced by a bug. The function build_image in Model/GAF_op.py returns images with shape 3×64×643 \times 64 \times 64 (channels first), but Model/GB.py:30 reads the first axis as the image height. The graph is then built on a 3×643 \times 64 grid, so all node coordinates lie between 0 and 2 on the vertical axis, as in the figure. I reproduced this bug by running the released code on a PeMS04 window.
Questions for authors
  1. Please describe the "region-to-sensor adapter" (Appendix C.2): its inputs, outputs and training, and how region features of different sensors are "matched" (lines 193-195). What is the MAE in Table 6 if the T-GRIG graphs are replaced by random graphs of the same size?
  2. Which model produces each row of Table 5, and what are the training settings? How is zero-shot transfer done between datasets with different numbers of sensors (Section 4.6)?
  3. Which data are used to find the morning peaks in lines 316-318, given that each input covers one hour? Are the pairs in Table 4 pairs of image regions or pairs of sensors, and how is the 5 km distance defined for them?
  4. Why are the 60-minute MAE values identical in Table 5 and in Tables 6-7 (19.47 on PeMS04 and 4.95 on METR-LA), although lines 270-272 state that these experiments use different inputs?
  5. On METR-LA, are zeros (missing values) excluded from MAE and RMSE? Please report results that can be compared with published ones (70:10:20 split, missing values excluded), and give the mean and standard deviation over several seeds.
  6. Which code produced Tables 2-8 and Figures 3-4, and will it be released?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Urgent Request for Program Chair and Area Chair Engagement Before Rebuttal Closure

Dear Program Chairs and Area Chairs,

Thank you for your work in coordinating the review of GB-Traffic. With the rebuttal discussion deadline approaching, we respectfully request your direct attention to the outstanding scientific discussion. We have submitted detailed, point-specific responses and followed up several times to invite technical engagement. Our latest explanations and controlled experimental results have yet to receive a substantive, point-specific assessment. The remaining discussion window makes timely coordination especially important.

The central research question is concrete and falsifiable. GB-Traffic evaluates three decisions: whether regional structural information improves traffic forecasting, which abstraction budget is most effective under a fixed predictor, and whether graph connectivity contributes predictive value when the representation size and relevant inputs are controlled. Section 4.3, Table 5, and Appendix U provide the corresponding matched comparisons and structural interventions.

The generalizability results directly evaluate how these decisions change across settings. The four-dataset analysis reveals distinct preferences. For example, medium regional input reduces BJMetro step-12 Ridge MAE from 138.5738 to 133.6607, while the corresponding comparison on METR-LA is 8.9929 versus 9.2269 in favor of the time-and-image input. Additional regular-partition comparisons, 100 independently trained residual MLP fits, 20 STBAN-1x32 fits, and directed cross-network tests examine construction, predictor, and network dependence. These experiments establish a repeatable diagnostic protocol and quantify the conditions under which different structural representations are useful.

At this stage, the discussion needs a direct assessment of these submitted results. We respectfully ask the Program Chairs and Area Chairs to review the specific author responses and encourage a focused follow-up identifying any remaining disputed experimental control, inference, or generalization dimension. This would allow us to address a concrete technical question before the discussion closes.

We also request that the final assessment explicitly consider the full rebuttal record. Concerns that have been answered with matched comparisons should be evaluated against those measurements, while any continuing objection should be connected to a specific scientific reason. The absence of a further exchange should not leave the submitted technical responses outside the final evaluation.

In summary, GB-Traffic offers an explicit benchmark question, controlled interventions, cross-dataset comparisons, predictor-dependent diagnostics, and cross-network evaluations. We have provided the corresponding experimental answers and repeatedly invited discussion. Given the approaching deadline, we respectfully request timely Program Chair and Area Chair engagement so that the final decision reflects these concrete results and any clearly articulated outstanding scientific concerns.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.