acceptodds
Withdrawn from ICLR 2027

FACO: Transferring Cross-City Learning Experience via Spatiotemporal Relay Function

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-city spatiotemporal forecasting aims to transfer predictive experience from data-rich cities to targets with limited observations. Existing methods typically assume continued access to source data, models, or computation, which may fail when historical sources and future targets are temporally separated. We study Forecasting spatiotemporal observations Across Cities Offline (FACO), where source-side learning is completed before the target becomes available and source-private observations, forecasting systems, and source-side computation are inaccessible during target adaptation. This setting raises three challenges: preserving predictive experience beyond source availability, maintaining consistent knowledge across heterogeneous cities and forecasting systems, and identifying trustworthy historical dynamics from limited target evidence. We develop FACO as an offline cross-city forecasting framework that aligns heterogeneous forecasting systems in predictive-functional coordinates, distills source-conditioned temporal evolution, preserves transferable dynamics through a frozen cross-city relay function, and adapts the relay to a future target using functional trust and closed-loop evolution. The relay is realized through a Spatiotemporal Relay Function (SRF), which enables source-derived dynamics to remain usable after source exit. Experiments across heterogeneous cities and forecasting architectures show that FACO achieves the strongest overall performance among the evaluated approaches, reducing MAE by 29.7% over STGP on Innsbruck under the three-day target adaptation setting. These results support representing cross-city predictive experience as source-independent dynamics that remain reusable beyond the availability and architecture of their source systems.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Cross-City Traffic Forecasting, Offline Knowledge Transfer, Few-Shot Adaptation;Heterogeneous Forecasting Systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_pg2qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FACO addresses cross-city traffic forecasting when source data and private forecasting systems become unavailable before target adaptation. It learns low-rank bilinear evolution functions, consolidates them into a frozen relay, and retains a shared interface initialization. A target adapts its interface using one or three days of observations and calibrates bounded, recursive forecast corrections. Evaluation covers 16 target cities, component ablations, augmentation of existing methods, and replacement of target backbones.

Critical strengths
  1. The deployment setting is practically relevant. A common predictive interface offers a sensible way to reuse knowledge across heterogeneous forecasting architectures without combining incompatible private-network parameters. Source validation, reliability weighting, and bounded corrections explicitly address potentially harmful transfer.

  2. The wrokload is very impressive, implementing numerous baselins for comparison.

Critical weaknesses
  1. The distinct contribution of the relay remains insufficiently isolated from conventional pretraining and additional modeling. Since this paper employ this kind of novel word to describe, but what's the difference on earth between adaptive learning and this work? It's just like pre-train-adaptive-learning. The “w/o Relay Evolution” ablation partially addresses this question; however, the paper should clarify whether it preserves the same pretrained interface and target adaptation procedure. Without this clarification, it is difficult to distinguish the relay’s contribution from the benefits of source pretraining and additional modeling capacity.

  2. Internal numerical discrepancies prevent a confident assessment of the reported advantages. Bolton's three-day FACO MAE/WAPE is 65.1/0.062 in Table 1 (p. 8), but the assigned-backbone reference is 87.01/0.0833 in Table 8 (p. 31). Also, under the same evaluation set and the aggregation definitions in Equations 91–92 (p. 27), MAE/WAPE must equal the mean absolute ground truth, independently of the predictor. Basel's TPB values, 168.4/0.346, imply approximately 486.7, whereas FACO's 136.9/0.284 imply approximately 482.0. This exceeds rounding precision. Different masks, aggregation conventions, or reporting errors could explain this; the current evidence does not identify the cause.

Questions for authors
  1. Clarification of existing results: Which configurations, splits, and runs produced the two Bolton entries? Please reconcile them and explain the Basel metric discrepancy, ideally providing sample counts, absolute-ground-truth sums, and absolute-error sums on a common evaluation set.

  2. Clarification and targeted evidence for attribution: Does “w/o Relay Evolution” preserve the same pretrained interface and target adaptation? Its numerical results could materially change my assessment. Additional comparisons against a target-only correction model of the same capacity and a conventional shared predictor with matched source data and comparable artifact/training budgets would more directly establish the relay's added value; these are proposed controls, not claims that such experiments are already reported.

Reviewer_T7pkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper looks at cross-city traffic forecasting when the data and models from earlier source cities are no longer available and proposes FACO. FACO stores what it learns from those cities in a frozen relay, then uses a small amount of data from a new city to decide how much of that knowledge to use. The paper then tests this proposed method on 16 target cities.

Critical strengths

S1: The paper studies a practically motivated setting in which target adaptation cannot access the original source systems.

S2: Code is released, which helps reproduction.

Critical weaknesses

W1: The paper equires joint optimization across historical source cities. But it is not clear when these source cities do not emerge in the same time, how training is performed.

W2: Semantic alignment relies on an element-wise affine transformation, which only adjusts feature scales and biases. It does not ensure that same semantic variables have consistent predictive effects across cities.

W3: Eq. (16) assigns source weights based on improvements over local predictors. However, within-source effectiveness does not imply cross-city transferability.

W4: Eq. (17) summarizes each source's functional support using only coordinate-wise means and standard deviations. overlooking multimodal distributions and feature correlations.

W5: 14 of 16 evaluated cities are located in Europe. This challenges its usage and generalization.

W6: WAPE is a constant multiple of MAE, yielding identical model rankings. Therefore, leading on both metrics does not provide independent evidence of superiority.

W7: Figure 3 reports ablation and hyperparameter sensitivity results for only 8 of the 16 target cities.

W8: Too many equations, making it hard to read.

Questions for authors

Please refer to weakness.

Reviewer_82hPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FACO studies cross-city traffic forecasting after historical source data, private forecasting systems, and source-side computation become unavailable. It constructs shared coordinates from normalized traffic histories, semantic context, and local predictive behavior. Before source exit, source ground-truth trajectories supervise bilinear temporal evolution functions and a shared interface initialization. Source functions are consolidated into a frozen relay with reliability and support summaries. A future target adapts the interface with one or three days of observations and calibrates bounded, trust-weighted recursive corrections. Evaluation covers 16 targets from an 18-city pool, three backbone families, component ablations, baseline augmentation, and hyperparameter sensitivity.

Critical strengths

S1. The deployment constraints are practically relevant, and normalized output-space coordinates offer a sensible mechanism for reuse across heterogeneous forecasting systems. Appendix D.3 and Table 8 test alternative target backbones while keeping source predictors and transferred memory fixed, providing relevant evidence for this design.

S2. The method explicitly controls source influence through held-out reliability, functional-support summaries, bounded corrections, and target-side relay calibration. The candidate set includes zero, allowing the target to select the uncorrected interface prediction. These mechanisms are useful design choices, although they do not guarantee improved test performance.

S3. The experimental coverage is broad. Tables 1 and 5 cover three-day and one-day adaptation on 16 cities; Tables 6 and 7 test baseline augmentation; Figure 3 includes component ablations; and Figures 6–11 examine adaptation and relay hyperparameters. The reported improvements are promising, subject to the consistency issues below.

Critical weaknesses

W1. The reported results require reconciliation. Bolton's three-day FACO MAE/WAPE is 65.1/0.062 in Table 1 (p. 8), but its assigned-backbone reference is 87.01/0.0833 in Table 8 (p. 31). The relationship between these runs is not explained. Also, Equations 91–92 (p. 27) imply that, on identical evaluation entries, MAE/WAPE equals the mean absolute ground truth and therefore cannot depend on the predictor. Basel's Table 1 values imply approximately 486.7 for TPB (168.4/0.346) and 482.0 for FACO (136.9/0.284); ordinary rounding at the displayed precision cannot reconcile them. Different masks or aggregation procedures could explain this, but would need disclosure. A further reporting mismatch concerns Bolton's one-day case study: Appendix D.5 and Figure 12 describe a full-test WAPE improvement of 12.9%, whereas Table 5's local-only/FACO values, 0.114/0.089, imply approximately 21.9%. These are unresolved reporting discrepancies, and they do not establish which underlying result is correct.

W2. The incremental contribution needs more precise attribution and positioning. The frozen artifact contains both a source-trained interface initialization ψ₀ and evolution memory M (Eq. 22). Figure 3 already provides relevant evidence through the “w/o Relay Evolution” variant, including relative MAE changes; the paper should not be treated as lacking this ablation. However, the exact retained components and training controls for that variant are not documented clearly enough. Please confirm whether it uses the identical ψ₀, target-interface adaptation, and model-selection procedure. A correction model of comparable capacity trained only on the permitted target prefix would further test whether transferring source dynamics adds value beyond additional correction capacity. Since FACO also retains a learned source-derived interface, its setting should explain more precisely which retained artifacts distinguish it from the pretraining approaches discussed in Appendix A. This is a request for clearer boundaries and attribution, rather than a conclusion that the method is equivalent to existing work.

W3. Some implementation details remain unresolved in the supplied PDF. Equation 37 and Appendix B.1 specify the predictive interface mainly by its inputs and outputs; the concrete architecture and its interaction with each backbone need an executable description. Appendix B.4 defines the agreement statistic and states how the agreement gate should behave, but does not give the explicit gate function. Appendix D.4 gives formulas for the support and horizon gates, which partially resolves the trust construction, but not the agreement gate. Appendix B.3 describes block resampling for source reliability without stating the block length, resampling count, or confidence level. Baseline descriptions in Appendix C.3 also need to specify how methods using source–target interaction are adapted to source unavailability. These omissions limit reproducibility from the PDF; I have not inspected any separate code supplement.

W4. Run-to-run uncertainty is not reported. Table 4 lists seed 42, and Appendix D.3 describes randomly assigned target backbone families. The presented results do not quantify variation across training seeds or repeated backbone assignments. This does not establish that the improvements are unstable, but it limits confidence in small margins such as Bolton's Table 1 MAE of 65.1 versus 66.4 for ST-MPPT. The shaded ranges in Figures 12–13 describe detector variation, not repeated-training uncertainty.

Questions for authors

Q1. Please reconcile the Bolton entries in Tables 1 and 8, the Basel MAE/WAPE ratios, and the Bolton one-day case-study percentage. For the metric check, please report the common evaluation-entry count, ground-truth absolute sum, and prediction-error absolute sum, and specify any method-dependent masks or aggregation.

Q2. What exactly is retained and retrained in “w/o Relay Evolution”? Does it preserve the same shared initialization and target adaptation? An absolute-error table alongside the existing relative-change plot would make the contribution easier to quantify. A matched target-only correction control would provide stronger attribution evidence.

Q3. Please specify the predictive-interface architecture, which backbone parameters are updated at each stage, the agreement-gate function, and the reliability-resampling settings. How are baselines requiring source–target interaction converted to the stated offline protocol?

Q4. Can the main comparisons be supplemented with repeated-seed results or a sensitivity check over backbone assignments, especially for the closest competitors?

Reviewer_V9XkOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors consider cross-city spatiotemporal forecasting when source cities and their forecasting systems are no longer available at the time of target adaptation. They propose FACO, which distills source-side predictive dynamics into a frozen relay function and adapts its use to a new city with limited local observations. The reported results across 16 target cities are strong. Nevertheless, the paper could more clearly establish which parts of the framework drive these gains and how the proposed offline setting affects the fairness of comparison with existing methods.

Critical strengths

1)The offline cross-city transfer setting addresses a meaningful deployment constraint. 2)The frozen relay offers a concrete way to reuse source-side knowledge after source exit.

Critical weaknesses

1)The framework contains several interacting components. The current ablations should more clearly quantify the contribution of each component across the target cities. 2)The paper should clarify whether every baseline is evaluated under exactly the same source-access restrictions at target adaptation time. Otherwise, the comparison may mix methods designed for different information settings. 3)The relay is frozen, but target-specific instantiation and trust calibration still require local observations. The method’s behavior when those observations are especially sparse or unrepresentative deserves further analysis. 4) The reported improvements are substantial, but the main results do not show run-to-run variation. Repeated-run statistics would help assess their robustness. 5) Some notation is unnecessarily unconventional or insufficiently explained, as exemplified by Eq. (6). The authors should use standard notation where possible and define each symbol clearly when it first appears.

Questions for authors

My questions for the authors are detailed in the weaknesses above.

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

This paper studies cross-city spatiotemporal forecasting when historical source cities and future target cities are temporally separated, and source data and forecasting systems are no longer accessible. In this setting, target cities can only use limited local observations and previously preserved knowledge for adaptation. The paper proposes FACO, which maps heterogeneous forecasting systems into shared predictive-functional coordinates, distills source-conditioned temporal evolution, and aggregates the learned knowledge into a frozen Spatiotemporal Relay Function. Functional Trust and closed-loop evolution are then used for target forecasting. Experiments across multiple cities and heterogeneous forecasting architectures show promising results.

Critical strengths
  1. The paper considers a practical scenario where historical collaborators leave, and source data and forecasting systems are no longer accessible. The problem is clearly defined and practically relevant.
  2. The paper compares FACO with multiple baselines across 16 target cities and provides ablation studies, hyperparameter sensitivity analyses, and case studies, offering relatively comprehensive experimental validation.
Critical weaknesses
  1. The method introduces many symbols, some of which are not clearly defined. The relationships between modules are also not intuitively explained, making the paper difficult to follow.
  2. The paper provides limited visualizations, particularly of the data representations and relationships between source and target cities.
  3. Several module designs lack sufficient motivation and intuitive explanation.
Questions for authors
  1. Could the authors clarify the specific implementations of the proposed modules, particularly the predictive interface and Adapt procedure? What are their corresponding backbones, inputs, outputs, and training procedures?
  2. The current experiments use 17 source cities for each target city. Could the authors evaluate different numbers and combinations of source and target cities, and analyze the relationships among source reliability weights, target-side Functional Trust, and final forecasting performance?
  3. Could the authors provide more direct qualitative analysis of whether the model learns transferable cross-city dynamics? For example, visualizing representations and evolution trajectories of different cities in the shared functional space, and analyzing the dynamics before and after transfer.

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

Related papers

Discussion (0)

Sign in to comment.