acceptodds
Under review as a conference paper at ICLR 2027

PGET-Net: Physics-Guided Neural Network with Event-Text Conditioning for Traffic Forecasting

Abstract

Accurate traffic forecasting during disruptive events is essential for dynamic traffic management and emergency decision-making, yet remains challenging. Disruptive events, such as extreme weather and traffic accidents, can disturb routine urban traffic patterns, thereby limiting the ability of methods that rely primarily on historical observations to accurately capture abrupt changes in traffic states. Although event-related text has been incorporated into forecasting models, it is often treated as auxiliary information rather than explicitly coupled with traffic dynamics. We propose PGET-Net, an event-text-conditioned physics-guided neural network that uses event semantics to modulate traffic-state evolution for multi-step forecasting. PGET-Net learns a nonlinear mapping from observed traffic states to a latent traffic potential field and models its evolution using diffusion-advection dynamics. The diffusion component describes routine propagation over the static road topology. The advection component maps event semantics to a graph representing event-impact propagation, which parameterizes event-induced perturbations to the latent dynamics. Both mechanisms are represented within a graph differential equation network, where a learnable gate adaptively combines their contributions to the evolution of the latent field. The resulting trajectory is decoded into future traffic prediction. Experiments on the Beijing Text-Traffic dataset (BjTT) show that PGET-Net outperforms the evaluated baselines, supporting the effectiveness of coupling event-conditioned perturbations with routine traffic propagation.

Primary area:
learning on time series and dynamical systems
Keywords:
Traffic Prediction, Physics-Guided Deep Learning, Emergency Events

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_tH6POpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PGET-Net that incorporates event-text information for traffic forecasting. Historical traffic states are represented through a latent traffic potential, whose evolution is motivated by diffusion and advection mechanisms. DiffGCN propagates on static route network, while AdvGCN is built on the graph generated by BERT encoded text and learnable node embeddings. The experiments are conducted on BjTT dataset.

Critical strengths
  1. The problem itself is well-motivated. Incorporating text event-information into traffic forecasting is challenging, especially for capturing traffic disturbances that are not reflected with historical measurement.
  2. The reported improvement during peak hour is significant.
Critical weaknesses
  1. The physical interpretation of the diffusion and advection mechanism is not convincing. Diffusion is described as the transport of traffic potential from higher-potential to lower-potential nodes. However, DiffGCN adopts a conventional GCN without establishing its relation to a diffusion operator or imposing constraints that enforce gradient-driven transport. In addition, it is unclear how this graph represents physical advection or directed traffic propagation with AdvGCN. In particular, the graph learned from textual information does not necessarily reflect road connectivity and traffic wave propagation. Therefore, the physical interpretation of the proposed framework is seriously questionable.

  2. The experiment comparison are insufficient to support the claimed contribution. The uniqueness of the BjTT dataset is ackonwledged. However, the selected baselines are relatively outdated, and none of them incorporate textual event information. Therefore, it is unclear whether the reported improvements arise from the proposed physics-guided design or simply from the additional information provided by event text.

  3. It is unclear how Eq. 6 connects to real traffic potential evolution and how the proposed equation relates to traffic flow propagation on road and network. A physically meaningful dynamical equation has to be considered. The proposed method is not technically sound.

Questions for authors

See weaknesses.

Reviewer_1iBROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PGET-Net, a physics-guided neural network for traffic forecasting under disruptive events. The framework maps historical traffic observations into a latent traffic potential field and models its evolution through diffusion and event-conditioned advection mechanisms within a graph differential equation network. Event text is encoded using BERT and cross-attention to construct an event-dependent propagation graph, while a learnable gate combines the diffusion and advection components. Experiments on the Beijing Text-Traffic dataset demonstrate improvements over several traffic forecasting baselines across multiple prediction horizons.

Critical strengths

S1. The paper studies a meaningful problem. Incorporating event semantics directly into latent traffic-state evolution sounds reasonable.

S2. The overall architecture is clearly outlined, and the role of each component is explained at a high level.

S3. The empirical study shows that the proposed method achieves the best performance across the evaluated prediction horizons on BjTT.

Critical weaknesses

W1. The major concern is that the comparison with baselines is not fair enough, since the authors explicitly state that PGET-Net additionally uses event text, but the baseline methods use only historical traffic data. Therefore, the performance improvements in Tables 1 and 2 cannot be attributed specifically to the proposed physics-guided architecture. It is difficult to isolate the benefit of the proposed event-to-dynamics coupling versus simply adding text information. It is suggested to add some event-aware baseline methods, such as FUSE-Traffic (FUSE-Traffic: Fusion of Unstructured and Structured data for Event-aware Traffic forecasting, SIGSPATIAL25)

W2. The ablation study in Table 3 compares different dynamics modules but does not include a variant without event text or an alternative text-aware method. To verify the contributions of the proposed method, a stronger ablation should include identical architectures with event text removed or mismatched event descriptions.

W3. The experimental study is only conducted on a single dataset. It would be better to use more datasets to show the generalization ability of the proposed method.

W4. The proposed PGET-NET is inspired by the Diffusion-Advection Equation. Equation (6) replaces the Laplacian and directional transport operators with two learned GNNs combined through an adaptive gate. The mathematical relationship between these GNN operators and the corresponding physical differential operators is not established. Therefore, the work is more suitable to be positioned as a physics-inspired method rather than a physically grounded traffic dynamics model.

W5. How are event texts spatially mapped to the network? How is the event-conditioned graph constructed? More such technical details should be provided to improve reproducibility.

Questions for authors

Please refer to the weakness.

Reviewer_3m2QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PGET-Net, a traffic forecasting model that incorporates disruptive-event text into the evolution of latent traffic states. First, PGET-Net encodes historical traffic observations into a latent traffic potential field. Then, it models the evolution of this field with a neural ODE with two components: (i) diffusion over the static road network and (ii) advection over an event-conditioned graph constructed from textual descriptions. A learnable gate combines these components, and a decoder produces multi-step traffic forecasts. Experiments on BjTT datasets demonstrates improvements over seven forecasting baselines.

Critical strengths
  1. Incorporation of event conditioning into dynamics as well as the input representations: PGET-Net constructs an event-conditioned graph that influences the evolution of latent traffic states. In my opinion, it seems that this modeling goes beyond standard feature-level multimodal fusion. It is reasonable to separate the fixed road geometry from event-dependent propagation.
  2. Consistent improvements across forecasting horizons, and promising ablation study results on components: Table 1 demonstrates the effectiveness of PGET-Net across all the three horizons. Table3 shows that the combined model outperforms both the diffusion-only and advection-only variants.
Critical weaknesses
  1. Inadequate isolation of the contribution of PGET-Net: All the seven baselines in Table1 use only historical traffic observations, whereas PGET-Net uses event text as well. This makes it difficult to interpret whether the improvement is due to the proposed diffusion-advection or using additional event text.
  2. Weak evidence for modeling disruptive-event propagation: This study conducts experimental evaluation that primarily consists of overall forecasting performance and a separate analysis of peak hours. However, peak-hour congestion is not necessarily a disruptive event, rather much of it follows regular daily patterns.
Questions for authors
  1. What event information is available in prediction?
  2. How better is PGET-Net than the baselines when the baselines use the same event text?
  3. Could you show direct evidence of the proposed event-induced propagation mechanism?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.