acceptodds
Under review as a conference paper at ICLR 2027

Learning Weekly Urban Evolution: Delay-Coupled Neural Dynamics for Extremely Long-Horizon Spatiotemporal Forecasting

Abstract

Understanding how cities evolve over long horizons is essential for capacity allocation, operational scheduling, and emergency management. As horizons extend from hours to days, city-wide traffic states evolve continuously across temporal phases, giving rise to complex spatiotemporal dynamics. Our analysis of real-world traffic data identifies three interrelated challenges: delayed interactions among nodes and their temporal accumulation, changing dependencies across phases and days, and the joint representation of long-term periodicity and local peak and trough variations. To address these challenges, we propose Delay-Coupled Neural Dynamics (DeCoDy), a graph dynamics framework with time delays for modeling extremely long-horizon urban evolution. DeCoDy addresses them through Multi-Delay Interaction and Accumulation, Phase Evolution and Adaptive Coupling, and Global and Local Residual Representation, while generating continuous future trajectories with a constrained full-horizon response solver. We evaluate DeCoDy on four real-world urban traffic datasets with horizons from 8 hours to 7 days. DeCoDy achieves the lowest MAE in 14 of 16 dataset–horizon settings and the lowest RMSE on all four datasets for 7-day forecasting, demonstrating its effectiveness in modeling extremely long-horizon urban evolution.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural dynamics, Spatio-temporal systems, Graph neural networks, Data mining

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1381OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose DeCoDy, a method for long-horizon traffic forecasting. They identify three challenges for long-horizon forecasting: (1) delayed interactions among nodes and their temporal accumulation, (2) changing dependencies across phases and days, and (3) the joint representation of long-term periodicity and local peak and trough variations. To address these challenges, the authors introduce a multi-delay interaction and accumulation module, a phase evolution and adaptive coupling module, and a global and local residual representation module. In experiments on four real-world datasets, DeCoDy outperforms the baseline methods.

Critical strengths
  • S1: On the four datasets, the proposed method achieves lower RMSE and MAE than the baselines in most cases, which empirically demonstrates its effectiveness compared to the baselines.
Critical weaknesses
  • W1: The authors raise three challenges for long-horizon forecasting, but the manuscript does not sufficiently explain why these challenges matter for long-horizon forecasting. This raises a fundamental question about the motivation of the work.
    • W1-1: In Figure 1(a), the drops at nodes C and D precede the drop at the target (node B) by only about 10 minutes. Such delayed interactions are short-term, and their temporal accumulation is not actually observed either. Moreover, there is no intuitive explanation of how these interactions could be exploited for multi-day forecasting.
    • W1-2: The paper does not explain why the continuous traffic evolution should pose a problem for long-horizon forecasting. In Figure 1(b), although the negative deviation persists for more than 100 steps across a day boundary, this is not a variation spanning the multi-day forecast horizons (288–2016 steps) that the paper targets.
  • W2: The presentation is unclear and difficult to follow.
    • W2-1: Some figures are difficult to understand. For example, in Figure 1(a), the axis labels do not describe the plotted quantities, so it is unclear what is being shown. In Figure 1(b), the manuscript does not explain what the vertical-axis label "deviation" refers to in the main body.
    • W2-2: In Section 3, many symbols are introduced without definition, which makes it difficult to follow the derivations and their meaning (e.g., \mu in Eq. (1), \phi in Eq. (3), \beta and \sigma in Eq. (4), \eta in Eq. (6), \rho in Eq. (7), ...).
Questions for authors
  • Q1: Related to W1, I understand that the three challenges raised in Introduction are all reasonable from the perspective of traffic forecasting, but could you explain why they become important problems specifically in long-horizon forecasting?
  • Q2: Figure 2 explicitly says "traffic flow." Is the proposed method restricted to traffic flow, with travel speed, density, or other traffic variables out of scope? In my understanding, "traffic flow" usually refers to the number of vehicles passing per unit time.
  • Q3: In the experiments, what is the intention behind changing the input sequence length depending on the forecast horizon?
Reviewer_XTcdOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DeCoDy, an urban evolution modeling method, which seperates periodical structure from date-specific deviations through the global and local residual representation module, captures evolving relationships through the phase evolution and adaptive coupling module, and models delayed interactions through the multi-delay interaction and accumulation modules. Empirical results on four real-world datasets show the effectiveness of the proposed method.

Critical strengths
  1. The paper is well-structured, presenting the method, especially the model components with clear logic, and perfectly matches the sections with the graphical illustrations. mathematical formulations explain the design step by step, providing plenty of detail.

  2. Comparison with recent and strong baselines demonstrates the advantage of the proposed methods.

Critical weaknesses
  1. Although the authors list some potential challenges to current traffic modeling methods, the paper lacks sufficient in-depth analysis to evaluate whether the observations are essential problems or just superficial noise, and it does not measure how much the models are affected. In the related work section, the authors claim that current methods fail to overcome these challenges. However, no representation-level analysis or case study is provided, making it hard to evaluate the extent to which they are confounded by the challenges. Hence, the motivation of this work lacks solid support.

  2. The model largely relies on extensive external information, such as the weather and the events, which might be unavailable or limited in many scenarios. They also involve high costs to collect and might be invalidated in the future. Given that the model structure is complicated, includes many components, and has high computational complexity, in case some external information is missing, wrong, or out-of-distribution, the delicate model is subject to performance degradation. In the experiments, the four datasets all fall in the US West Coast background, sharing high similarity in weather and events. The results lack validation under different natural climates and regional cultures, and there are no domain transfer or adaptation experiments.

Questions for authors

How is the calendar, weather, and event information represented, stored, and encoded in the model? Please specify.

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

This paper focuses on the delay and extreme-long (days to week-level) traffic forecasting by proposing a new predictor called DeCoDy. The authors conducted experiments on several open benchmark datasets.

Critical strengths
  • S1: Overall the draft is prepared with care and made understandable.
Critical weaknesses
  • W1: This work studies a well-studied problem, i.e., traffic forecasting, by extending to an extremely long horizon. However, I'm not convinced by the meaning of this extension. Daily to weekly level traffic forecasting heavily depends on the data (how many anomalies/OOD you get): it's gonna be either trivially predictable by periodicity or dominated by extrinsic events, for which you have an encoder but there is not description/not your focus. Thus, this work appears to be really incremental in comparison to the existing ones, especially given you're comparing to some new preprints doing the long-horizon forecast.

  • W2: The motivations are somehow nebulous too. They look having been considered and modeled by existing works, e.g., PDFormer [1]

[1] Pdformer: Propagation delay- aware dynamic long-range transformer for traffic flow prediction. AAAI 2023

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.