acceptodds
Withdrawn from ICLR 2027

Transferable Spatio-Temporal Node Identities: Elevating Multilayer Perceptrons to Leading Cross-City Zero-Shot Traffic Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Spatio-temporal forecasting models are conventionally trained and deployed per city, and the learnable node embeddings that empower state-of-the-art in-domain predictors are bound to the training network, precluding zero-shot transfer to unseen cities. We propose Transferable Spatio-Temporal Node Identities (TransSTNI), a paradigm that reconstructs node identity as a function computed from deployment-time priors rather than retrieved from dataset-bound parameters. TransSTNI derives node identity from three transferable sources: the Graph-Conditioned Structural Identity (GSI) encodes each node's structural role from the spectral coordinates of its own topology via a permutation-equivariant encoder; the Window-Conditioned Temporal-State Identity (WTI) captures node-level temporal states through multi-reference statistics computed over the current history window; and the Road-Attribute Semantic Embedding (RASE) provides functional context from de-geolocated road attributes through a frozen text encoder. These representations condition a residual MLP shared across all nodes, whose parameter count is independent of network size. To prevent memorization of the source topology, source-domain training interleaves full-graph learning with episodic optimization on variable-size induced subgraphs. At deployment, all parameters remain frozen: GSI and RASE are computed once and cached, and each window is predicted by a single forward pass without target-domain labels or adaptation. Experiments on LargeST demonstrate that TransSTNI achieves state-of-the-art zero-shot performance, while retaining the inference efficiency.

Primary area:
learning on time series and dynamical systems
Keywords:
zero-shot, traffic forecasting

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_a2pAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cross-city zero-shot traffic forecasting, motivated by the fact that conventional learnable node embeddings are tied to source-city node indices and cannot be directly transferred to unseen graphs. The proposed TransSTNI framework replaces fixed node identities with computable ones: GSI derives structural identity from the target graph topology, WTI extracts node-level temporal states from statistics of the current history window, and RASE encodes transferable road attributes. These representations condition a shared residual MLP, while variable-size subgraph training is introduced to reduce source-topology memorization. Experiments on the SD, GBA, and GLA regions of LargeST evaluate six zero-shot transfer directions, together with efficiency, sensitivity, and ablation analyses.

Critical strengths

The paper identifies a real limitation of identity-based traffic forecasting: node-indexed embeddings become undefined when the node set or topology changes. TransSTNI replaces these fixed parameters with identities computed from target topology, current temporal context, and road attributes, allowing the same predictor to operate on networks of different sizes. This is well aligned with the frozen-parameter zero-shot setting, and the six LargeST transfer directions provide encouraging empirical support.

Critical weaknesses

W1. GSI first recomputes Laplacian eigenvectors F(A)F(A) from the input graph, while Eq. 8 only shows equivariance of the subsequent encoder assuming the node features are already transformed as PFPF. However, Laplacian eigenvectors have sign ambiguity, and repeated or near-repeated eigenvalues further introduce eigenspace basis ambiguity. Therefore, the paper does not establish that the full mapping A→F(A)→ZGSIA \rightarrow F(A) \rightarrow Z^{GSI} is permutation-equivariant as claimed. Since GSI is a core part of the transferable identity design, this affects the technical justification of the method rather than being a minor presentation issue.

W2. Table 1 includes STUNet, but several recent generalizable or foundation baselines are missing, such as OpenCity, UrbanDiT, UniFlow, UrbanFM, and FactoST-v2. Moreover, all zero-shot transfers are conducted among three regions from the same LargeST benchmark, which share the same data source and preprocessing pipeline. The current evidence therefore supports cross-region transfer within LargeST, but is less conclusive for broader unseen-city generalization and the SOTA claim.

[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. How does GSI handle the sign ambiguity of Laplacian eigenvectors and the basis ambiguity associated with repeated or near-repeated eigenvalues? This directly affects the claimed permutation-equivariance of the full GSI pipeline.

Q2. Can the authors clarify or provide comparisons with recent generalizable or foundation methods such as OpenCity, UrbanDiT, UniFlow, UrbanFM, and FactoST-v2? If these methods are not directly comparable because of different pretraining resources or task protocols, the paper should clearly explain the mismatch.

Q3. Can the authors provide a zero-shot transfer experiment on a dataset outside LargeST?

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

The paper introduces TransSTNI, a lightweight framework for zero-shot cross-city traffic forecasting that replaces fixed node-indexed embeddings with transferable, input-conditioned node representations. The proposed architecture integrates graph-conditioned structural identities derived from Laplacian spectral features and GraphSAGE, window-conditioned temporal identities constructed from multi-reference traffic statistics, and road-attribute semantic embeddings obtained using a frozen pretrained text encoder. These representations condition a shared residual MLP forecasting backbone. To improve generalization to unseen road networks, the authors introduce a two-stage training procedure combining full-graph supervised learning with variable-size induced-subgraph optimization.

Experiments on three LargeST regional datasets, covering six directed zero-shot transfer scenarios, report consistently lower MAE, RMSE, and MAPE than the evaluated baselines. Additional hyperparameter studies, component ablations, supervised comparisons, and runtime experiments support the claims regarding the utility of the representation modules and the computational efficiency of the framework.

Critical strengths

1. Strong empirical performance in zero-shot cross-city traffic forecasting.

The paper demonstrates consistently strong forecasting performance across six cross-city transfer scenarios on the LargeST dataset, outperforming the evaluated baselines in MAE, RMSE, and MAPE. These improvements are achieved without target-domain fine-tuning or parameter adaptation, despite differences in sensor counts and road-network topologies. This provides meaningful empirical evidence for the practical effectiveness of the proposed framework in addressing cross-network generalization, a challenging problem in spatio-temporal forecasting.

2. Efficient and transferable node representation framework.

The paper presents a practical approach to overcoming the limitations of dataset-specific node embeddings by constructing node identities from graph structure, historical temporal statistics, and road attributes. Integrating these representations with a shared residual MLP enables forecasting on previously unseen sensor networks while maintaining low computational overhead. The reported runtime comparisons demonstrate substantial efficiency improvements over existing methods, and the ablation experiments provide evidence for the contributions of the proposed representation modules, particularly WTI.

Critical weaknesses

1. Insufficient validation of the proposed transferable representations.

Although the paper reports promising zero-shot forecasting results, the experimental analysis does not sufficiently establish why the proposed representations generalize across cities. In particular, the ablation study only evaluates the removal of entire modules, without examining the individual statistical components of WTI or the effectiveness of the spectral features used in GSI. Furthermore, the claimed permutation equivariance of GSI is not fully justified, as the formulation does not address the sign and basis ambiguities of Laplacian eigenvectors.

2. Concerns regarding experimental fairness and the strength of the zero-shot comparisons.

The authors modify certain baselines by removing node-specific spatial embeddings to enable cross-city evaluation, but do not sufficiently justify whether these modifications preserve their effectiveness or represent competitive zero-shot implementations.

3. Incomplete comparison with relevant zero-shot forecasting approaches.

The experimental evaluation does not consider LLM-based or pretrained foundation-model approaches for spatio-temporal forecasting such as ST-LLM+, despite their demonstrated zero-shot prediction capabilities. Their omission limits the strength of the paper's claims regarding state-of-the-art zero-shot forecasting performance and the accuracy–efficiency advantages of the proposed MLP-based architecture.

Questions for authors
  1. How does GSI handle the sign and basis ambiguities of Laplacian eigenvectors across different graphs? Does the permutation-equivariance claim apply to the complete GSI module?

  2. Which baselines were modified for zero-shot evaluation, and how was their competitiveness ensured?

  3. How were the ablation results in Table 5 aggregated? Why do the full-model results differ from the corresponding averages in Table 1? Can the authors report variability across multiple runs?

  4. How much do GSI, RASE, and Stage 2 training independently improve generalization? How does the framework compare with a simpler temporal-statistics-based MLP?

  5. What is the exact Stage 2 optimization procedure, particularly regarding the reported meta-learning parameters? How is its contribution to cross-graph generalization established?

Reviewer_Tt1AOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes TransSTNI, a model for zero-shot cross-city traffic forecasting: the model is trained on the sensors of one city and then used in another city without any training there. The model is based on STID, a residual MLP that predicts each sensor from its recent history, time-of-day and day-of-week embeddings, and a learnable embedding of each sensor. TransSTNI replaces the learnable sensor embeddings with three inputs that can be computed for a graph with any number of sensors: GSI, the Laplacian eigenvectors of the road graph passed through GraphSAGE; WTI, statistics of the current 12-step input window of each sensor, such as its mean, trend, and percentile rank among all sensors; and RASE, a Qwen3-Embedding vector of a text built from the road metadata of each sensor. Training has two stages: first on the full graph of the source city, then also on smaller graphs, each built from a random 50-75% of the source sensors and the edges between them. The evaluation uses three regions of LargeST (SD, GBA, GLA) with 15-minute data from 2019, a 60:20:20 split, and 12 input and 12 output steps. When trained on one region and tested on another, TransSTNI has the lowest errors for all six pairs of regions (Table 1) and the shortest training and inference times (Table 2).

Critical strengths

The problem is practical: a model that needs no training data from a new city can forecast traffic there from the first hours after the sensors are installed. Replacing the learnable sensor embeddings of STID with inputs computed from the graph, the input window, and the road metadata is a simple way to run one model on graphs of any size, and the resulting model is small and fast.

Critical weaknesses

1. The paper is hard to read.

Much of the text uses abstract and decorated wording instead of simple statements. For example, the abstract describes TransSTNI as "a paradigm that reconstructs node identity as a function computed from deployment-time priors rather than retrieved from dataset-bound parameters" (lines 17-20), which means only that the sensor embeddings are computed from the graph, the input window, and the road metadata instead of being learned for each sensor.

2. The experiments do not show that the proposed components work.

  • In the ablation study (Table 5), only removing WTI increases the error substantially (MAE 44.03 vs 33.34 on SD). Removing GSI, RASE, or Stage 2 increases the MAE by 0.01 to 1.48, and in three of nine cases by at most 0.05: 33.39 vs 33.34 without GSI on SD, 31.91 vs 31.90 without RASE on GLA, and 31.95 vs 31.90 without Stage 2 on GLA. Differences this small can come from random variation between training runs, and all results come from a single run with one seed (stage2/oastid_full.py:2383), yet the paper describes "a clear performance drop" when RASE is removed (lines 743-744). Moreover, RASE encodes four simple metadata fields (road class, direction, number of lanes, detector type) with Qwen3-Embedding, but the paper does not compare it with a one-hot encoding of the same fields, so there is no evidence that the text encoder is needed. Finally, in the code, the variant without GSI is also trained without Stage 2 (submit_egbeta_limnbr_repro.sh:24,32), although the paper states that all other settings are kept "identical to the full model" (lines 722-723).
  • Laplacian eigenvectors are defined only up to their sign, and the code does not fix the signs (stage2/oastid_full.py:93-113), which may explain why removing GSI barely changes the results.
  • When trained and tested on the same region, TransSTNI reaches an MAE of 25.62, 27.11, and 26.76 on SD, GBA, and GLA (Table 4), which the paper calls a "fully supervised upper bound" (Appendix A.4). Under the same LargeST setting, GWNET reaches 17.74, 20.91, and 21.20, and LSTM, which uses no data from other sensors, reaches 26.44, 27.96, and 28.05 [1]. Therefore, TransSTNI trained on the target region is only slightly better than LSTM. Moreover, its zero-shot MAE of 30.84-34.66 in Table 1 is much higher than the MAE of standard models trained on the target region, and the paper reports no such results for the baselines in Table 1.
  • All three regions come from one California dataset, and the paper compares with none of the zero-shot foundation models it cites, such as OpenCity and UrbanFM. Therefore, the claim that this work is "the first to realize cross-city zero-shot forecasting with fine-grained node discrimination" (lines 476-477) is not supported.

3. The zero-shot results use test data of the target city.

The paper states that the method "only uses target topology and current historical windows" (lines 142-143). However:

  • The script that PACKAGE.md:19 names as the main entry point normalizes the target city with one mean and one standard deviation computed over its whole 2019 series, including the test period (lib/dataloader.py:35-36,91-96, called from stage2/oastid_full.py:948-950). The predictions are converted back to vehicle counts with the same two numbers. A second entry script normalizes each input window separately (RevIN), which the paper never mentions. It is unclear which of the two scripts produced Table 1.
  • The hyperparameters were chosen on the target test sets: Figure 2 shows "Test set MAE and RMSE" (line 396), the text says that this figure confirms "the optimal configuration" (lines 430-431), and the script that computes the plotted errors uses the target test sets (scripts/eval_ckpt_rmse_bm.py:41-44). Table 1 reports results on the same test sets.

4. Several numbers are copied from another paper or do not match each other.

  • In Table 2, the training and inference times of nine baselines (DCRNN, AGCRN, STGCN, GWNET, ASTGCN, STGODE, DSTAGNN, DGCRN, D2STGNN) are identical to Table 3 of the LargeST paper [1], which measured them on an RTX A6000 GPU, while TransSTNI was timed on an A800 GPU (lines 651-652). The paper does not say that these numbers are taken from LargeST or that they were measured on a different GPU. The caption still refers to "four datasets" (line 433), as in LargeST, although the table has three, and the paper does not cite LargeST at all.
  • In Table 1, ST-GFSL has exactly the same MAE, RMSE, and MAPE for both source regions of each target region, for example, 47.21, 67.08, and 49.79% for both models tested on GBA, one trained on SD and one trained on GLA. Two models trained on different regions are very unlikely to give the same three numbers. Moreover, ST-GFSL is a few-shot method: its original code, included in the repository, fine-tunes the model on three days of target-city data. The paper does not say how ST-GFSL was run without target-city data.
  • The paper does not say which source and target regions each column of the ablation study (Table 5) uses. The SD column equals the average of the two results in Table 1 for training on SD and testing on GBA or GLA, so each column seems to be a source region averaged over its two targets. However, the same average for GLA gives an MAE of 30.97, while Table 5 reports 31.90.

[1] Liu et al. "LargeST: A Benchmark Dataset for Large-Scale Traffic Forecasting," 2023

Questions for authors
  1. Which normalization was used for the target city in Table 1: one mean and standard deviation computed over the whole target series, or a separate normalization of each input window? How do the results change when no statistics of the target series are used?
  2. How do the results change when all hyperparameters are chosen on the validation set of the source city, and what are the standard deviations over several seeds?
  3. How does TransSTNI compare with zero-shot foundation models such as OpenCity?
  4. The best zero-shot MAE in Table 1 is 30.84, for training on GLA and testing on GBA. According to the LargeST paper, models trained on GBA itself reach a much lower MAE: 20.91 for GWNET and 27.96 for LSTM. Sensors in a new city record data from the day they are installed. How many days of these data do GWNET or LSTM need to reach a lower MAE than zero-shot TransSTNI? If they need only a few days, the practical motivation for zero-shot forecasting in this setting is weak.

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

Related papers

Discussion (0)

Sign in to comment.