acceptodds
Under review as a conference paper at ICLR 2027

What Transfers Across Cities? Invariant Representation Learning for Robust Spatio-Temporal Forecasting

Abstract

Cross-city forecasting requires transferring knowledge across urban environments that share predictive structure yet differ in traffic dynamics, observation density, and latent environmental factors. Many existing approaches align source and target distributions, but alignment can also retain source-specific correlations and induce negative transfer when domain mismatch becomes substantial. We introduce CauCrossNet, an invariant representation transfer framework that separates relatively stable predictive structure from transient environment-specific variation and promotes prediction consistency across alternative stable-space summaries. The learned invariant space provides a basis for target adaptation. Target-side augmentation is conditioned on stable representations, while source knowledge is integrated according to representation-level transfer reliability rather than unconditional similarity. This design is causality-inspired but does not assume an identifiable structural causal model or require interventional observations. Across multiple city pairs, we evaluate standard transfer, limited-target-data adaptation, temporal shifts, and longer horizons. Results show that selective invariant transfer improves forecasting across source-target relations, remains effective under mismatch and limited target data, and reduces degradation from poorly matched source information through reliability-aware integration. These findings suggest that robust cross-city forecasting depends not only on reducing domain discrepancy, but also on identifying predictive structure that remains stable, useful, and transferable across environments. Code and experimental configurations are available at the anonymous repository: https://anonymous.4open.science/r/CauCrossNet-040B.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Cross-City Forecasting, Invariant Representation Learning, Transfer Learning, Spatio-Temporal Forecasting, Distribution Shift

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Lcf6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies negative transfer in cross-city spatio-temporal forecasting and asks what information should actually be transferred when source and target cities differ substantially. It proposes CauCrossNet, which decomposes representations into a relatively stable branch ZIZ_I and an environment-specific branch ZEZ_E. CauKD learns the transferable representation through consistency regularization and MMD-based alignment, CauCL performs invariant-conditioned target augmentation and curriculum learning, and RAFusion controls source contribution according to representation mismatch and predictive uncertainty. The evaluation covers multiple transfer directions, few-shot adaptation, temporal shifts, cross-region OOD transfer, and longer forecasting horizons. The paper explicitly frames its contribution as predictive invariant transfer rather than causal identification.

Critical strengths

The paper correctly points out that reducing source-target discrepancy does not necessarily imply that the aligned information is transferable, since source-specific correlations may still induce negative transfer. The method performs well not only under standard transfer, but also under few-shot, cross-region OOD, temporal-shift, and long-horizon settings. For example, on PeMS03→PeMS08, the average MAE decreases from 17.55 for the strongest baseline D2MHyper to 14.64 for CauCrossNet. This makes the empirical result more convincing than a gain limited to a single benchmark setting.

Critical weaknesses

W1. The experiments mainly compare against earlier cross-city transfer methods such as DASTNet, DAGN, CCL, D2MHyper, MTPB, and ST-GFSL, while recent models such as OpenCity, UrbanDiT, UniFlow, UrbanFM, and FactoST-v2 are not systematically discussed or evaluated. The paper currently discusses PLM-CUP as a recent complementary method, but the broader positioning against recent spatio-temporal foundation or generalization models remains limited. Although some of these methods may use different data formats or adaptation protocols and therefore may not admit a direct numerical comparison, the paper should clearly explain these incompatibilities and evaluate those that can be adapted to the graph-based sensor setting. Without this, it is difficult to assess the current empirical competitiveness and novelty of CauCrossNet relative to recent cross-city spatio-temporal modeling.

W2. The roles of ZIZ_I and ZEZ_E are induced by the training objectives, but the paper does not directly quantify whether ZIZ_I is indeed more stable across environments than ZEZ_E, nor whether RAFusion weights systematically track actual source usefulness or negative-transfer risk. The authors themselves acknowledge that there is no direct quantitative ZIZ_I vs ZEZ_E role-identification test and that RAFusion is not independently evaluated as a calibrated uncertainty mechanism. Thus, the evidence is stronger for “the full pipeline works” than for the proposed mechanistic interpretation of why it works.

[1] Li et al., Open Spatio-Temporal Foundation Models for Traffic Prediction, ACM TIST, 2025.

[2] Yuan et al., UrbanDiT: A Foundation Model for Open-World Urban Spatio-Temporal Learning, ICLR, 2025.

[3] Yuan et al., UniFlow: A Foundation Model for Unified Urban Spatio-Temporal Flow Prediction, IEEE TMC, 2026.

[4] Chen et al., UrbanFM: Scaling Urban Spatio-Temporal Foundation Models, KDD, 2026.

[5] Zhong et al., Learning to Factorize and Adapt: A Versatile Approach Toward Universal Spatio-Temporal Foundation Models, 2026.

Questions for authors

Q1. Could the authors provide a more systematic comparison or positioning against recent methods such as OpenCity, UrbanDiT, UniFlow, UrbanFM, and FactoST-v2? For methods that cannot be directly compared, please clarify whether the incompatibility comes from the data representation, forecasting protocol, or adaptation setting.

Q2. Could the authors provide more direct diagnostics of the roles of ZIZ_I and ZEZ_E, and of the behavior of the RAFusion gate? For example, comparing the two branches in terms of cross-city or workday/non-workday discrepancy, their standalone predictive utility, or showing how source weights change with increasing mismatch would help establish that the proposed mechanism behaves as intended.

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

This paper proposes CauCrossNet, an invariant representation transfer framework for robust cross-city spatio-temporal forecasting. The method aims to mitigate negative transfer by separating stable predictive representations from environment-specific variations. It consists of three components: CauKD for invariant representation learning, CauCL for invariant-conditioned target augmentation, and RAFusion for reliability-aware source-target prediction fusion. Experiments on traffic flow and speed datasets demonstrate promising improvements over existing baselines across multiple transfer settings.

Critical strengths
  1. The paper addresses the important problem of negative transfer in cross-city spatio-temporal forecasting. The motivation for selectively transferring stable predictive knowledge rather than aligning entire feature distributions is meaningful and well justified.

  2. The overall framework is coherent and well motivated. Combining invariant representation learning, target-side augmentation, and reliability-aware fusion provides a systematic approach to handling cross-city distribution shifts and limited target data.

  3. The evaluation covers multiple traffic flow and speed datasets, four standard transfer directions, and additional settings involving limited target data, temporal shifts, and longer prediction horizons. The reported results are promising and demonstrate the potential effectiveness of the proposed framework.

Critical weaknesses
  1. The invariance of the learned representations is not directly validated. The method separates features into invariant and environment-specific representations and applies MMD alignment to the invariant branch. However, minimizing distribution discrepancy does not necessarily guarantee that the learned representations capture stable and transferable predictive relationships.

  2. The relationship between predictive uncertainty and transfer reliability is insufficiently justified. RAFusion reduces the contribution of source knowledge when predictive uncertainty increases. However, high predictive uncertainty does not necessarily imply that source knowledge is unreliable. In some cases, a target node with high uncertainty may benefit particularly from transferable source patterns.

  3. The effectiveness of reliability-aware fusion in identifying beneficial and harmful transfer is not directly demonstrated. Although the ablation results support the usefulness of RAFusion, they do not establish whether the learned fusion weights accurately reflect the actual reliability of source knowledge.

Questions for authors

Please refer to the weaknesses.

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

This paper studies cross-city traffic forecasting and proposes CauCrossNet, which transfers only the predictive structure that stays stable across cities instead of aligning full distributions. It separates stable and transient representations, aligns only the stable part, conditions target augmentation on it, and weights source predictions by representation mismatch and uncertainty. CauCrossNet outperforms 19 baselines in four transfer directions on PeMS, METR-LA, and PeMS-BAY.

Critical strengths
  1. The paper addresses negative transfer in cross-city forecasting. I find the distinction between what to transfer and how much to align particularly useful.
  2. The evaluation covers several settings, including few-shot learning, cross-region transfer, temporal shifts, and long-horizon forecasting. CauCrossNet outperforms the 19 baselines across all four transfer directions.
Critical weaknesses
  1. The central claim, that aligning only the stable branch ZIZ_I beats ordinary alignment, remains untested, because there is no same-backbone comparison with full-representation alignment (acknowledged in l. 472–474).
  2. All five datasets are small, with only 170–358 sensors and 2–5 months of data each, so it is unclear whether the method scales to large benchmarks such as LargeST (Liu et al., 2023) or TraffiDent (Gou et al., 2025).

References Liu et al. (2023). LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting. NeurIPS. Gou et al. (2025). TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents. NeurIPS.

Questions for authors
  1. Could you compare aligning only ZIZ_I with aligning the full representation, using the same backbone?
  2. Could you report results on a larger benchmark, for example transfer between LargeST (Liu et al., 2023) regions (e.g., GBA→SD)?
  3. The GBRT row in Table 12 is identical to the VAR row in Table 10. Could you check this and clarify how the baseline results were obtained?

Comments on the paper

2 comments
AuthorsOpenReview ↗

Request for Review Leadership Attention to the Completed Experimental Responses Before the Discussion Deadline

Dear Program Chairs and Area Chairs,

We sincerely appreciate your efforts in coordinating a rigorous and constructive review process for our submission, What Transfers Across Cities? Invariant Representation Learning for Robust Spatio-Temporal Forecasting (Paper 16000).

As the discussion deadline approaches, we respectfully request your attention to the substantial additional experiments and technical responses submitted during the rebuttal period, particularly those addressing the remaining concerns of Reviewers ujnG and fZJV.

First, we substantially expanded the forecasting evaluation. We evaluated five contemporary methods across four cross-city transfer directions, covering 48 forecasting comparisons. CauCrossNet achieves 36 strictly best results and one jointly best result, including leading results across all 24 comparisons in the two traffic-flow transfer directions.

Second, we directly addressed Reviewer ujnG's concerns about large-scale forecasting and baseline reliability. Our additional experiments include real-world LargeST forecasting in both GBA→SD and SD→GBA directions, as well as three-seed evaluation on the 2,409-sensor Orange County network in TraffiDent. Furthermore, we traced the duplicated VAR/GBRT values to Table 3 of a published Knowledge-Based Systems paper and provided independent rerun results for the affected configurations. The complete numerical results and experimental explanations have been presented in our latest responses.

Third, we provided two dedicated technical replies to Reviewer fZJV's follow-up questions. These responses include three-seed stable/specific representation diagnostics, equal-capacity predictive probes, a mathematical explanation of the near-zero routing weights, and controlled comparisons between learned and fixed source-admission coefficients. We explicitly explain the fixed-0.5 result and distinguish source-benefit ordering from fusion-weight calibration.

Additionally, nine matched alignment experiments directly compare stable-only alignment, full-representation alignment, and the no-MMD configuration using the same backbone.

We are grateful that Reviewer Lcf6 explicitly acknowledged that the additional experiments improved their overall assessment and indicated an increase in rating.

Meanwhile, the other two reviewers' current ratings remain at 2. Since their latest questions have motivated further substantial experimental work and detailed responses, we respectfully request that the review leadership consider these new materials and encourage a final technical reassessment of the outstanding concerns before the discussion closes.

We would particularly appreciate your guidance on whether any specific scientific questions still require clarification. We remain fully available to respond promptly and provide precise explanations.

Our aim is to ensure that the final assessment reflects the complete scientific record, including the original submission, the expanded experimental evaluation, and the detailed exchanges during rebuttal.

Thank you again for your time, careful consideration, and commitment to a comprehensive evaluation of our work.

Best regards,

The Authors

AuthorsOpenReview ↗

Final Author Statement: Scientific Contribution, Rebuttal Results, and Holistic Assessment

Dear Area Chair, Senior Area Chairs, and Program Chairs,

As the discussion concludes, we respectfully ask that Paper 16000, What Transfers Across Cities? Invariant Representation Learning for Robust Spatio-Temporal Forecasting, be assessed on its scientific contribution and the complete rebuttal record, beyond the arithmetic mean of review scores.

1. A distinctive scientific contribution. CauCrossNet addresses a central problem in cross-city forecasting: which predictive information remains useful when urban environments change? Its contribution is the coordinated use of a stable representation to guide (i) cross-city alignment through CauKD, (ii) target-side augmentation through CauCL, and (iii) source-knowledge admission through RAFusion, while retaining city-specific predictive information. This integration distinguishes our approach from conventional distribution alignment and shared/private decomposition.

2. Strong forecasting results. Across the four original transfer directions, CauCrossNet achieves average MAEs of 14.64, 2.03, 20.33, and 1.26, corresponding to improvements of 16.58%, 15.77%, 5.97%, and 17.65% over the strongest retained published comparators. Our expanded comparison includes OpenCity, UrbanDiT, UniFlow, UrbanFM, and FactoST-v2. Across 48 reported metric cells, CauCrossNet ranks strictly first in 36 and jointly first in one, including 24/24 first-place results in the two traffic-flow directions.

3. A substantive response to each major scientific question. During rebuttal, we presented nine same-backbone alignment runs comparing stable-only, full-representation, and no-MMD conditions. We added three-seed stable/specific branch diagnostics, equal-capacity forecasting probes, decoded gate interventions, item-level source-benefit analyses, and uncertainty-stratified results. We explicitly examined why a fixed 0.5 gate attains lower average MAE than the learned gate in the reported intervention, and explained the role of the subsequent source-conditioned refinement.

We also supplied real-data forecasting results for both LargeST directions (GBA→SD and SD→GBA) and TraffiDent transfer to 2,409 target sensors. For the repeated VAR/GBRT entries, we traced the values to the cited Knowledge-Based Systems table, explained their inclusion, and supplied independent reruns on the affected targets. These replies directly address the outstanding questions with numerical results, protocols, and technical explanations.

4. The timing of the record matters. Reviewer Lcf6 explicitly acknowledged an improved assessment and indicated a higher rating. Reviewer ujnG's November 18 assessment preceded our additional LargeST, TraffiDent, and baseline-reexecution replies submitted later that day. Reviewer fZJV's latest questions also received two dedicated technical replies on November 19. We ask that these later responses be evaluated on their merits, with the chronology taken into account.

5. Request to the review leadership. The final decision should weigh novelty, significance, forecasting accuracy, methodological analysis, and the full author–reviewer exchange. Review scores summarize individual judgments at particular moments; their numerical average cannot substitute for an independent assessment of the complete scientific record.

We respectfully ask the Area Chair and Program Chairs to consider the substance of each objection and our latest response, rather than relying on the score average alone. We believe CauCrossNet presents a strong case for acceptance at ICLR 2027, and we appreciate your careful, holistic judgment.

Thank you for your consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.