acceptodds
Withdrawn from ICLR 2027

Towards Efficient, Scalable, and Transferable Spatio-Temporal Forecasting: A Periodic Perspective

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Transferable spatio-temporal forecasting seeks to reuse knowledge from data-rich regions when target observations are scarce. Recurring patterns offer transferable information, yet their signal levels, fluctuation amplitudes, and short-term evolution vary across regions. We propose PACT, a graph-free forecasting framework that uses recent observations and future daily-cycle positions to guide the application of shared temporal regularities. PACT predicts local changes relative to each node's latest observation in a normalized space, then refines these predictions using information aggregated across nodes. Its core mechanism conditions a compact global summary on both the future cycle position and the current regional state, enabling node-specific retrieval of context for each forecast step. Shared parameters accommodate different node sets without adjacency matrices or cross-region node correspondence. The compact aggregation and retrieval operations avoid pairwise node interactions, yielding computation that scales linearly with node count at fixed model dimensions. Experiments on real-world traffic datasets demonstrate strong forecasting performance across regions and sensing scales, together with efficient target adaptation and inference. Code is available at https://anonymous.4open.science/r/PACT-F67B/.

Primary area:
learning on time series and dynamical systems
Keywords:
Cross-City Traffic Forecasting, Spatio-Temporal Few-Shot Learning, Periodic Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_U6VCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a transferable, graph-free framework for spatio-temporal forecasting. The key idea is to use shared periodic patterns across different regions through conditional global aggregation. The framework is evaluated on cross-city traffic forecasting tasks and shows consistent improvements over baselines.

Critical strengths
  1. The evaluation results shows consistent improvements over baselines.
  2. Avoids explicit graph construction and pairwise node interactions. Reduce computational complexity.
Critical weaknesses
  1. The paper does not characterize the periodic similarity or heterogeneity between source and target datasets, which makes it difficult to understand how these properties affect transfer performance.
  2. The experimental baselines do not include more recent transferable forecasting methods.
Questions for authors
  1. Could the authors provide a quantitative analysis of the periodic similarity between different cities?
  2. The proposed framework relies on shared periodic patterns across regions. However, it is unclear whether the framework remains effective when the source and target regions exhibit different periodic patterns. Could the authors investigate how transfer performance varies with the degree of periodic similarity between regions? This would help clarify the role of shared periodic patterns in the transferability.
  3. The most recent experimental baselines are from 2024. Could the authors provide comparisons with any recent methods? What are the key advantages of PACT over more recent approaches?
Reviewer_dhvzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a transferable spatiotemporal prediction framework PACT, which reduces the impact of signal level differences through relative change encoding, and uses future daily cycle positions and current regional states to jointly adjust global aggregation, and then corrects local predictions through node queries. The model does not require adjacency matrices or cross regional node correspondences, and the computational complexity increases linearly with the number of nodes in a fixed model dimension.

Critical strengths
  1. Parameters are shared between nodes without the need for adjacency matrices or node correspondences, and can directly adapt to regions with different numbers of nodes, reducing the cost of structural adaptation during migration.
  2. In addition to module ablation, the paper examines the role of components through interventions such as cycle shift, state replacement, and removal of global corrections, and analyzes the concentration and stability of aggregation weights, which helps to understand model behavior.
Critical weaknesses
  1. The appendix discusses SOFTS using aggregation and distribution mechanisms, as well as OpenCity for cross city prediction, but it was not included in the experiment
  2. After removing the status condition, the MAE only increased from 2.88 to 2.92 within 60 minutes. The appendix also found that some constant condition interventions can improve LA → BAY, and the stable benefits of combined state conditions still need to be verified.
  3. The one-day adaptation of LA → BAY in Table 6 is actually inferior to the zero sample, indicating that the adaptation effect may be significantly dependent on the composition of the support set.
Questions for authors

See Weaknesses

Reviewer_m67qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PACT, a graph-free framework for cross-city traffic forecasting with limited target data. It predicts each node’s change from its latest observation and refines it using a compact global summary conditioned on the future time of day and the current city state. Across eight transfer tasks, PACT achieves the best MAE and MAPE in all settings, and the paper also includes ablation, zero/few-shot and efficiency analyses.

Critical strengths
  1. Simple, small and fast. It needs no road graph or sensor matching between cities, wins in every transfer setting in the paper, and is clearly the fastest. That is useful for a new city with little data, and it is what I like most about this paper.
  2. The motivation makes sense. The same recent trend can lead to different futures depending on the time of day, and the authors show the model really uses this time signal.
Critical weaknesses
  1. It is unclear whether the new design is what makes PACT better. PACT starts from the latest reading and adds a small change, which is already hard to beat for short-term traffic. Even with no data from the new city, it beats the best baselines trained on two days in most cases, and its lead mostly shrinks at longer horizons.
  2. The novelty is moderate. Starting from the latest reading, pooling all sensors into a summary and using time of day are common tricks. Letting time of day guide the summary is a sensible addition, but the ablation does not fully separate its effect: removing time of day altogether mainly shows that time information helps.
Questions for authors
  1. How does PACT compare with a simple baseline that just repeats the latest reading?
  2. Do the baselines also receive time of day as input? If not, it would help to see one or two strong baselines given the same starting point and time-of-day input as PACT.
  3. METR-LA stores missing readings as zeros, which seem to be included in the MAE. Could you clarify how they are handled, and report MAE with them excluded?
  4. If time of day is given directly to each sensor’s own prediction instead of being used to build the city-wide summary, does the model still do as well? This would show whether the new design itself is needed.

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

Related papers

Discussion (0)

Sign in to comment.