acceptodds
Under review as a conference paper at ICLR 2027

Physics-Informed Directed Spatio-Temporal Graph Convolutional Networks for Traffic Flow Prediction

Abstract

Accurate urban traffic flow prediction is crucial for Intelligent Transportation Systems (ITS). Existing spatio-temporal neural networks typically adopt symmetric adjacency matrices, failing to capture the asymmetric spatial dependencies of traffic flow, and lack macroscopic physical constraints, leading to non-conservative predictions in long-term forecasting. To address these issues, we propose PI-Dir-STGCN. Unlike prior methods that loosely couple physical penalties with graph networks, our framework achieves a tight structural coupling between directed graph structures and macroscopic traffic dynamics. Specifically, we construct an asymmetric directed weighted adjacency matrix to capture these asymmetric traffic dependencies, and develop a graph-topology generalized upwind difference scheme to embed the LWR model into the graph neural network. This innovation transforms the directed graph into a numerical grid for fluid dynamics and encourages mass-conserving predictions through a computable physics residual. Experiments on PeMS04 and PeMS08 show that our model effectively suppresses cumulative errors, integrates data-driven learning with physical evolution, and provides strong physical interpretability.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Traffic flow prediction, Directed spatio-temporal graph convolution, Physics-informed deep learning, Lighthill-Whitham-Richards model, Mass conservation regularization

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_9RXXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes PI-Dir-STGCN for traffic flow forecasting. The method combines directed graph convolution with a physics-inspired conservation constraint based on the Lighthill–Whitham–Richards traffic flow model. It also predicts traffic flow, occupancy, and speed jointly, with the goal of improving forecasting accuracy while encouraging physically consistent predictions.

Critical strengths
  1. Incorporating traffic flow conservation into graph-based forecasting is a reasonable direction. The use of road direction information is also relevant to the physical characteristics of traffic networks.

  2. The paper includes several ablation experiments, such as STGCN-Multi, Dir-only, and PI-only, which help examine the contributions of the additional inputs and individual components. Reporting conservation error is also useful.

Critical weaknesses
  1. The forecasting performance is not competitive with the stronger baselines. In Table 1, the proposed method obtains MAE values of 23.36 and 18.73 on PeMS04 and PeMS08, respectively, while STAEformer achieves 18.23 and 13.58. These are substantial differences. Moreover, Table 3 shows that the improvement over STGCN-Multi is only 0.31 and 0.32 MAE on the two datasets. This raises questions about the practical value of the proposed physics-informed and directed-graph components.

  2. The physical formulation needs stronger justification, As I see, the paper uses occupancy as a proxy for traffic density and introduces routing weights to describe traffic flow between connected nodes, however, it is unclear whether the resulting graph-based discretization consistently represents the underlying conservation law, particularly at intersections involving merging or diverging traffic streams.

  3. The experimental evaluation is limited to two datasets, it gives the relatively small improvement over the matched STGCN-Multi baseline, more extensive experiments would be necessary to establish that the proposed physical constraints offer consistent benefits.

Questions for authors
  1. Can the authors provide a more rigorous derivation of the graph-based conservation constraint, including its treatment of road intersections, boundary conditions, and physical units?

  2. Why does the proposed method perform substantially worse than STAEformer on both datasets? What is the practical advantage of PI-Dir-STGCN compared with stronger existing forecasting models?

  3. Can the authors evaluate the proposed method on additional datasets and provide results across multiple random seeds, particularly for the comparisons with STGCN-Multi?

3 comments
AuthorsOpenReview ↗

We provide a rigorous mathematical reconstruction, verify conservation via a minimal network, and demonstrate practical superiority through multi-seed and plug-and-play experiments.

We deeply appreciate your meticulous review and constructive feedback. R1:We have thoroughly rewritten Section 3.2.3 and Appendix A.1 to address your concerns.Units: We introduce a learnable node-specific scalar c_i to convert occupancy to density, and a global scalar α to dynamically absorb dimensional discrepancies. Intersections and Boundaries: We explicitly define the closed-boundary assumption and clarify that remaining residuals stem from unmonitored ramps. Furthermore, we designed a minimal closed-network verification experiment (new Section 4.6). A 4-node network with diverging/merging flows proves that our routing ratios sum to exactly 1.0, guaranteeing structural flux conservation (internal residual = 0.0) at arbitrary intersections. R2: We acknowledge in the revised text that STAEformer (MAE 18.23/13.58) is the most accurate pure data-driven point estimator. However, pure data-driven models suffer from severe physical inconsistency (high CE) and error accumulation. The practical advantage of PI-Dir-STGCN is its physical reliability and long-horizon stability. More importantly, our proposed physical module is universal: we have integrated our LWR residual into a multivariate STAEformer (STAEformer-Multi) to demonstrate the universality of our method, and presented this in 4.2. Specifically, after incorporating the LWR loss, the MAE on PeMS04 dropped from 20.80 to 19.73, and on PeMS08 it decreased from 15.17 to 14.88. Meanwhile, under a unified evaluation metric (using unscaled, inverse-transformed raw predictions without α), the Conservation Error (CE) also demonstrated a consistent decline (from 0.135 to 0.112 on PeMS04, and from 0.101 to 0.097 on PeMS08).We clearly maintain in Table 1 that the original STAEformer (yielding MAEs of 18.23 and 13.58) remains the most accurate pure data-driven point estimator. Our newly added paired experiment is specifically designed to answer what benefits are gained when applying the physical loss to the exact same multivariate backbone: it acts as an effective plug-and-play regularizer that not only improves macroscopic physical consistency but also enhances point-prediction accuracy by mitigating overfitting. R3:We have updated Table 3 to report the mean and standard deviation across 3 random seeds (42, 1024, 2026) to ensure rigorous comparison, particularly against STGCN-Multi. While time constraints during the rebuttal period prevent us from evaluating entirely new datasets, we believe the current comprehensive evaluation on two major benchmarks (PeMS04 and PeMS08) with multi-seed runs sufficiently validates our claims.

Area_Chair_pfqpOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 9RXX,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_9RXXOpenReview ↗

Thanks for the rebuttal

Thank you for the rebuttal. After reviewing the concerns raised by the other reviewers, I have decided to maintain my original score.

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

The paper offers limited novelty. Its main contribution appear to only modify STGCN into a directed adjacent graph and add a loss based on macroscopic LWR model. Neither modification represents a methodological innovation. More importantly, the experiment results on PeMS04/08 are worse than SOTA (23.38 vs. 18.23 of STAEformer in MAE on PeMS04, and 18.73 vs. 13.58 of STAEformer on PeMS08). Meanwhile, the selected baselines are largely outdated, making it difficult to establish the novelty of the proposed approach. Given the limited technical contribution, weak experimental result, and inadequate baseline comparisons, it raises concern about the scientific rigor and originality of the work.

Critical strengths

Authors attempt to incorporate LWR traffic flow model with GNN that improves interpretability of the approach.

Critical weaknesses
  1. The paper lacks methodological novelty. STGNN-based methods have been extensively studied for eight years and subsequent works have already incorporated directed graphs and more advanced methods have further explored heterogeneous graphs to capture heterogeneity in traffic. Moreover, the LWR physics-informed component is only a soft penalty in the loss function, rather than being integrated into the underlying mechanism. Existing PINN present as provide more rigorous and well-motivated approaches.
  2. The experimental performance is weak. The proposed method fails to achieve SOTA performance and is worse than STAEformer, a baseline published in 2023. Insufficient discussion of clear advantage compared with STAEformer.
Questions for authors

see weakness

2 comments
AuthorsOpenReview ↗

We honestly clarify the "soft regularizer" nature, redefine our methodological novelty based on forward-flux routing, and demonstrate the plug-and-play value on the STAEformer backbone.

We appreciate your critical and insightful comments regarding our methodological novelty and experimental performance. R1:We agree that directed graphs and STGNNs are not new, and we honestly acknowledge in our revision that the LWR component acts as a soft regularizer rather than a hard underlying mechanism. However, our specific novelty lies in how to distribute traffic flux along the actual driving direction and how to construct a computable discrete residual based on this specific flux routing. We introduce a forward-routing ratio ω_ji to strictly isolate forward advection (as verified by our new minimal-network experiment in Section 4.6). We have replaced abstract claims like "tight structural coupling" with precise mathematical descriptions of the computational alignment between our directed graph topology and the generalized upwind scheme. R2:We acknowledge that STAEformer achieves stronger point-prediction accuracy through heavy parameterization. To show the clear practical advantage of our method, we do not compete on absolute MAE alone, but demonstrate that our physics module can serve as a plug-and-play regularizer for STAEformer. Adding our LWR residual to STAEformer simultaneously improves its MAE (e.g., from 20.80 to 19.73 on PeMS04) and reduces its Conservation Error. This proves the practical value of our method in building reliable, physically consistent models for long-horizon forecasting.This is added to Section 4.2.

Area_Chair_pfqpOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer beVi,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_18LuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PI-Dir-STGCN, a physics-informed directed spatio-temporal graph convolutional network for traffic flow forecasting. The authors identify three gaps in existing work: (1) most STGNNs use undirected or symmetric graph convolutions that violate traffic flow directionality and lack macroscopic dynamic constraints; (2) existing physics-informed methods rely on 1D road discretization that does not generalize to arbitrary graph topologies; and (3) neither data-driven nor physics-informed approaches effectively suppress error accumulation in long-horizon multi-step forecasting. To address these, the paper makes three contributions: an asymmetric directed weighted adjacency matrix with random-walk normalized directed graph convolution; a graph-topology generalized upwind discretization that aligns message passing with the advection operator of the LWR PDE; and a physics residual regularization term (plus a flux-consistency loss) added to the training objective. Experiments are conducted on PeMS04 and PeMS08.

Critical strengths
  1. The directed-graph upwind discretization is a principled and non-trivial coupling of graph message passing with the LWR conservation law.

  2. The long-horizon error analysis provides evidence for the central claim about error accumulation.

Critical weaknesses
  1. The empirical comparison is not convincing relative to the stated state of the art, and the strongest baseline undermines the paper's framing. Table 1 shows STAEformer achieves substantially lower errors than the proposed model on both datasets (PeMS04: MAE 18.23 vs. 23.36; PeMS08: 13.58 vs. 18.73; RMSE 30.30 vs. 36.18 and 23.43 vs. 28.91). The authors dismiss this by arguing STAEformer is a "deterministic black box restricted to single-variable prediction" that "precludes mathematical verification of mass conservation." This is a weak defense: physical interpretability is a desirable property, but the paper's own quantitative results show a 22–28% MAE gap to a model the authors themselves label state-of-the-art. A physics-informed model that is markedly less accurate than a non-physics baseline does not establish that the physics prior is a net win for the forecasting task; at best it establishes a trade-off, which the paper does not characterize rigorously

  2. The novelty relative to existing physics-informed graph traffic forecasting is not adequately positioned. The paper cites Pan et al. (2025) on physics-guided STGNN for intersection traffic but does not compare against it or any other physics-informed graph baseline; the only physics-aware comparison is an internal ablation (PI-only vs. Dir-only). Meanwhile, most of the modules in the proposed methods are have been largely studied. The uniqueness of the proposed methods are not well demonstrated.

Questions for authors
  1. Accuracy–conservation trade-off: STAEformer outperforms your model by 22–28% MAE on both datasets. Can you provide a Pareto analysis (accuracy vs. Conservation Error) that includes STAEformer and, ideally, a physics-regularized variant of a stronger backbone (e.g., STAEformer + your LWR residual)? Without this, how should a reader weigh the accuracy gap against the interpretability benefit?

2.Conservation verification: The closed-mainline assumption s^t_i≈0 is violated by on/off-ramps in PeMS. Can you quantify cumulative mass drift over the 60-minute rollout for your model versus STGCN-Multi, and report the fraction of nodes with ramps in each dataset?

  1. Novelty positioning: How does your graph-topology generalized upwind scheme differ from existing finite-volume/finite-element discretizations on graphs used in physics-informed GNNs outside traffic (e.g., in fluid or climate modeling)? For example, a direct technical comparison against at least one external physics-informed graph baseline (e.g., Pan et al., 2025).
3 comments
AuthorsOpenReview ↗

We provide a paired experiment (STAEformer + LWR) and Minimal-Network Verification

We sincerely thank you for these constructive comments. We address your three questions separately below. R1:Following your excellent suggestion, we have integrated our LWR residual into a multivariate STAEformer (STAEformer-Multi) to demonstrate the universality of our method, and presented this in 4.2. Specifically, after incorporating the LWR loss, the MAE on PeMS04 dropped from 20.80 to 19.73, and on PeMS08 it decreased from 15.17 to 14.88. Meanwhile, under a unified evaluation metric (using unscaled, inverse-transformed raw predictions without α), the Conservation Error (CE) also demonstrated a consistent decline (from 0.135 to 0.112 on PeMS04, and from 0.101 to 0.097 on PeMS08).We clearly maintain in Table 1 that the original STAEformer (yielding MAEs of 18.23 and 13.58) remains the most accurate pure data-driven point estimator. Our newly added paired experiment is specifically designed to answer what benefits are gained when applying the physical loss to the exact same multivariate backbone: it acts as an effective plug-and-play regularizer that not only improves macroscopic physical consistency but also enhances point-prediction accuracy by mitigating overfitting. R2:Thank you for this insightful question. To distinguish between conservation guaranteed by our algorithm and missing information in real-world data, we designed a minimal closed-network verification experiment (4-node network with diverges and merges). Under steady boundary inflow, the routing ratios strictly sum to 1.0, and the flux residual at internal nodes equals 0.0. This mathematically shows our graph operator preserves physical conservation within the observable topology. The residuals observed on PeMS largely stem from unmeasured on- and off-ramps. Faced with such imperfect data, our model acts as a robust physical regularizer that encourages observable links to follow the conservation law. We have added this in the new Section 4.6. R3:Thank you for pointing out the need to clarify our exact novelty. Our graph-topology generalized upwind scheme differs fundamentally from existing finite-volume/finite-element discretizations used in general physics-informed GNNs (e.g., fluid or climate modeling). We highlight two key distinctions: 1.Compared to general physics-informed GNNs: These methods typically solve PDEs on fixed, isotropic meshes where spatial coupling is symmetric. Our key advancement is adapting the upwind difference scheme to a directed, non-Euclidean graph by explicitly decoupling the message-passing graph (Matrix A, which allows backward shockwave propagation) from the physical forward-routing flux graph (Matrix B and routing ratio ω). This ensures the discrete spatial derivative inherently respects one-way traffic advection. 2.Compared to Pan et al. (2025): While Pan et al. focus on queuing dynamics and signal timing optimization at signalized intersections, our framework is fundamentally designed for segment control volumes and continuous flux advection in arterial and freeway networks.

Area_Chair_pfqpOpenReview ↗

Please Respond to the Authors’ Rebuttal

Dear Reviewer 18Lu,

The discussion period ends in a few hours. Could you kindly check the authors’ responses and let us know whether your concerns have been addressed?

Thank you!

Best regards, The AC

Reviewer_18LuOpenReview ↗

Thanks for the rebuttal

Dear Authors and AC,

Thanks for the rebuttal work and the extras experiments. It addesses some of my concerns. The states about the novelty of the paper is more clear. Considering the topic and the key contribution of this paper and the exsting similar paper. I would keep my score as 2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.