Bridge: Retrieval-Augmented Spatiotemporal Modeling for Urban Delivery Demand

Dingyi Zhuang; Junlin He; Lijun Sun; Qianjun Huang; Tong Nie; Yihong Tang

arxiv: 2605.19172 · v1 · pith:UVKTFQ6Enew · submitted 2026-05-18 · 💻 cs.LG · cs.AI

Bridge: Retrieval-Augmented Spatiotemporal Modeling for Urban Delivery Demand

Yihong Tang , Tong Nie , Junlin He , Qianjun Huang , Dingyi Zhuang , Lijun Sun This is my paper

Pith reviewed 2026-05-20 11:39 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords urban delivery demandcold-start forecastingretrieval-augmented modelingspatiotemporal graphscross-city transferdemand predictiongraph neural networks

0 comments

The pith

Bridge retrieves matching past region-time patterns to improve delivery demand forecasts in areas lacking any history.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses forecasting urban delivery demand in newly added service regions that have no historical records. Standard spatiotemporal models can capture spatial patterns once data accumulates but stay parametric and miss short-term operational behaviors in these cold-start cases. Bridge adds a time-aware memory store of region-time windows and retrieves relevant future demand trajectories for each target region by matching both its context and recent dynamics. A gated fusion step then combines the retrieved patterns with the output of an inductive graph backbone. The retriever itself is trained with a future-aware loss so that selected memory entries improve actual forecast accuracy rather than mere similarity.

Core claim

Bridge augments an inductive contextual graph forecaster with a retrieval mechanism over a memory of prior region-time windows. For any target region the model pulls future demand sequences whose regional context and recent dynamics align with the query, then refines the backbone prediction through gated fusion. Training the retriever with a future-aware objective ensures that retrieved entries are chosen for their utility in forecasting rather than surface similarity alone.

What carries the argument

Time-aware memory of region-time windows retrieved by joint regional context and recent dynamics, then combined via gated fusion and optimized by a future-aware objective.

If this is right

Resource allocation and routing in newly launched delivery zones can rely on shorter data collection periods.
Cross-city model deployment becomes practical even when only partial observations are available in the target city.
Operational planners gain an explicit memory of comparable past situations rather than depending solely on learned parameter generalization.
Forecast reliability improves in regions whose land-use or temporal rhythms resemble previously observed areas.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same retrieval-plus-graph pattern could extend to other cold-start spatiotemporal tasks such as traffic speed or energy load prediction.
Hybrid systems that keep a non-parametric memory alongside parametric models may become standard for operational forecasting where new sites appear regularly.
Memory size and retrieval granularity become new hyperparameters that trade off storage cost against forecast quality in large-scale deployments.

Load-bearing premise

That pulling future demand patterns from a memory of similar past region-time windows can recover short-term operational dynamics that a parametric graph model cannot learn from context alone.

What would settle it

An ablation experiment on the four delivery datasets that removes the retrieval and fusion components entirely and measures whether accuracy in within-city cold-start and cross-city transfer settings drops to the level of the plain graph backbone.

Figures

Figures reproduced from arXiv: 2605.19172 by Dingyi Zhuang, Junlin He, Lijun Sun, Qianjun Huang, Tong Nie, Yihong Tang.

**Figure 2.** Figure 2: Overview of BRIDGE. Region contexts and historical demand are encoded for inductive graph forecasting; a time-aware memory retrieves future priors from similar region-time patterns; the retrieved priors are then fused with the graph forecast for final prediction. regions Rinact t = RO t \ Ract t as inactive, whose histories are masked to simulate cold-start conditions. During training, we construct an indu… view at source ↗

**Figure 3.** Figure 3: Ablation of the Lret in the cross-city transfer setting. Table II summarizes the cross-city transfer results over all 12 ordered source-target pairs. BRIDGE achieves the best MAE, RMSE, and R2 in every pair, outperforming the graphonly backbone as well as MTGNN, IGNNK, and STGCN. The gains over the graph-only backbone are sometimes modest, but they are consistent: the average MAE decreases from 3.904 to 3… view at source ↗

**Figure 4.** Figure 4: Qualitative forecasting examples. Left column: predicted and ground-truth demand curves for representative regions in Hangzhou and Chongqing. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

read the original abstract

Forecasting urban delivery demand becomes substantially more challenging when newly added service regions lack historical records. Existing spatiotemporal forecasters effectively model spatial dependence once sufficient node histories are available. Still, they remain parametric and therefore struggle to recover short-term operational dynamics in cold-start regions. Geospatial embeddings help identify where a region is and what function it serves, yet they do not directly reveal how a similar region behaves under a comparable temporal context. We propose Bridge, a retrieval-augmented spatiotemporal graph framework that combines an inductive contextual graph backbone with a time-aware memory of region-time windows. For each target region, Bridge retrieves future demand patterns from the memory using both regional context and recent dynamics, and refines the backbone forecast through a gated fusion mechanism. To align retrieval with forecasting utility, we further train the retriever with a future-aware objective that favors entries whose future trajectories best match the target. Experiments on four real-world delivery datasets show that Bridge consistently improves over competitive spatiotemporal baselines in both within-city cold-start and cross-city transfer with partial observations. The results show that retrieval augmentation provides a useful operational memory for cold-start urban demand forecasting when parametric graph generalization alone is insufficient.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Bridge adds retrieval from a time-aware memory to an inductive graph for cold-start urban delivery forecasting, with plausible but thinly supported gains and a training objective that needs closer scrutiny for leakage.

read the letter

The main takeaway is that Bridge pairs an inductive spatiotemporal graph with a retrieval step that pulls future demand patterns from a memory of past region-time windows, using regional context plus recent dynamics and a future-aware training objective to pick the right matches. It then fuses the retrieved signal back into the backbone forecast via a gate. This targets the practical problem where new delivery zones have no history and parametric models cannot recover short-term operational patterns from embeddings alone. The abstract reports steady improvements over baselines on four real delivery datasets for both within-city cold-start and cross-city transfer settings. That combination of retrieval augmentation and the specific future-aware objective is new relative to prior graph or retrieval work in this area, and the framing around operational memory for logistics is a useful angle. The idea itself is straightforward and addresses a gap that practitioners actually face. The experimental claims rest on consistent outperformance, which is the kind of result that could matter for deployment. On the downside, the reported results give no error bars, no significance tests, and no clear account of temporal split rules or data exclusion. That makes it difficult to gauge how reliable the gains are. The future-aware retriever objective also raises the exact leakage issue flagged in the stress test: if memory entries are scored or selected using trajectory information that overlaps the evaluation horizon, the advantage over pure parametric baselines could be partly artifactual rather than a clean demonstration of context-driven recovery. A reader would want to see the precise masking logic in the methods. This work is aimed at people doing spatiotemporal forecasting for mobility and logistics who already know the standard graph baselines. Someone facing cold-start regions in delivery or similar domains would find the architecture worth trying out. It has enough of a concrete proposal and application focus to merit a serious referee, even though the current evidence is preliminary. I would send it to review with specific requests for ablations on the retrieval objective, full statistical reporting, and explicit checks against temporal leakage.

Referee Report

2 major / 2 minor

Summary. The paper proposes Bridge, a retrieval-augmented spatiotemporal graph framework for urban delivery demand forecasting in cold-start regions lacking historical records. It combines an inductive contextual graph backbone with a time-aware memory of region-time windows; for each target, it retrieves future demand patterns using regional context and recent dynamics, then refines the backbone forecast via gated fusion. The retriever is trained with a future-aware objective that favors memory entries whose future trajectories match the target. Experiments on four real-world delivery datasets report consistent improvements over competitive spatiotemporal baselines in within-city cold-start and cross-city transfer with partial observations.

Significance. If the central experimental claims hold under stricter validation, the work offers a practical way to augment parametric spatiotemporal models with non-parametric retrieval for data-scarce urban forecasting tasks. The combination of inductive graph modeling and time-aware memory retrieval addresses a real operational gap, and the future-aware objective is a notable design element that could generalize to other cold-start prediction settings. The results, if robust, would strengthen the case for retrieval augmentation when pure graph generalization is insufficient.

major comments (2)

[Abstract / Proposed method] Abstract and proposed method section: The future-aware retriever objective favors memory entries whose future trajectories best match the target. For within-city cold-start regions, this risks indirect leakage if temporal splits do not strictly prevent overlap between training retrieval scoring and the test horizon; please provide a precise description of how the memory bank construction and similarity computation enforce isolation from future data in the evaluation windows.
[Experiments] Experiments section: The abstract states consistent improvements on four datasets yet provides no details on error bars, statistical significance tests, data exclusion rules, or post-hoc selection criteria. These omissions make it impossible to assess whether the reported gains over baselines are robust or could be artifacts of experimental choices; full tables with variance, p-values, and protocol details are required.

minor comments (2)

[Proposed method] Clarify the exact definition of 'regional context' and 'recent dynamics' used for retrieval scoring; the current description is high-level and could benefit from a concrete formulation or pseudocode.
[Proposed method] The gated fusion mechanism is mentioned but its implementation details (e.g., gating function, training stability) are not elaborated; add a short description or equation reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on temporal isolation and experimental reporting. We address both major comments below with clarifications and commit to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

Referee: [Abstract / Proposed method] Abstract and proposed method section: The future-aware retriever objective favors memory entries whose future trajectories best match the target. For within-city cold-start regions, this risks indirect leakage if temporal splits do not strictly prevent overlap between training retrieval scoring and the test horizon; please provide a precise description of how the memory bank construction and similarity computation enforce isolation from future data in the evaluation windows.

Authors: We appreciate the referee's attention to this critical detail. The memory bank is populated exclusively from the training temporal window (all region-time entries with timestamps strictly before the validation and test periods). During training of the retriever, the future-aware objective computes similarity only against future trajectories that lie within the same training window, ensuring no test-horizon data influences scoring. At inference for cold-start regions, retrieval uses only the target's observed recent dynamics up to the current time step, with no access to any future observations. We will add an explicit subsection (Section 3.4) detailing the temporal partitioning, memory construction protocol, and similarity computation to make this isolation unambiguous. revision: yes
Referee: [Experiments] Experiments section: The abstract states consistent improvements on four datasets yet provides no details on error bars, statistical significance tests, data exclusion rules, or post-hoc selection criteria. These omissions make it impossible to assess whether the reported gains over baselines are robust or could be artifacts of experimental choices; full tables with variance, p-values, and protocol details are required.

Authors: We agree that the current experimental presentation lacks sufficient statistical rigor. In the revised version we will expand the Experiments section with: (i) error bars showing standard deviation across five independent runs with different random seeds; (ii) paired statistical significance tests (t-tests with Bonferroni correction) reporting p-values for all main comparisons; (iii) explicit data exclusion rules (regions with fewer than 30 observations or excessive missing values are removed); and (iv) a complete protocol appendix describing hyper-parameter search, early-stopping criteria, and any post-hoc analyses. Updated result tables will include these details for all four datasets. revision: yes

Circularity Check

0 steps flagged

No significant circularity: retrieval augmentation and future-aware objective are independent architectural components.

full rationale

The paper's derivation introduces an inductive contextual graph backbone augmented by a separate time-aware memory bank and retriever. The future-aware training objective aligns retrievals to forecasting utility by favoring matching future trajectories, but this is a standard supervised auxiliary loss applied during training on regions with history; it does not redefine the target demand forecast in terms of itself or reduce the final gated-fusion prediction to quantities fitted within the same equations. No self-citations are invoked as load-bearing uniqueness theorems, no ansatz is smuggled, and no known empirical pattern is merely renamed. The central claim rests on the proposed architecture's ability to recover dynamics in cold-start settings, which is externally falsifiable via the reported experiments on four datasets rather than tautological by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that similar regions exhibit transferable demand dynamics under comparable contexts; no explicit free parameters or invented entities are named in the abstract.

axioms (1)

domain assumption Similar regions exhibit similar demand patterns under comparable temporal contexts.
Invoked to justify retrieval from memory using regional context and recent dynamics (abstract description of Bridge).

pith-pipeline@v0.9.0 · 5746 in / 1219 out tokens · 36523 ms · 2026-05-20T11:39:54.064642+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We propose BRIDGE, a retrieval-augmented spatiotemporal graph framework that combines an inductive contextual graph backbone with a time-aware memory of region-time windows... train the retriever with a future-aware objective that favors entries whose future trajectories best match the target.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Experiments on four real-world delivery datasets show that Bridge consistently improves over competitive spatiotemporal baselines in both within-city cold-start and cross-city transfer

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 4 internal anchors

[1]

Effects of urban delivery restric- tions on traffic movements,

G. Yannis, J. Golias, and C. Antoniou, “Effects of urban delivery restric- tions on traffic movements,”Transportation Planning and Technology, vol. 29, no. 4, pp. 295–311, 2006. 1

work page 2006
[2]

Real-time demand forecast- ing for an urban delivery platform,

A. Hess, S. Spinler, and M. Winkenbach, “Real-time demand forecast- ing for an urban delivery platform,”Transportation Research Part E: Logistics and Transportation Review, vol. 145, p. 102147, 2021. 1, 2

work page 2021
[3]

Autonomous robot- driven deliveries: A review of recent developments and future direc- tions,

S. Srinivas, S. Ramachandiran, and S. Rajendran, “Autonomous robot- driven deliveries: A review of recent developments and future direc- tions,”Transportation research part E: logistics and transportation review, vol. 165, p. 102834, 2022. 1

work page 2022
[4]

Joint estimation and prediction of city-wide delivery demand: A large language model empowered graph-based learning approach,

T. Nie, J. He, Y . Mei, G. Qin, G. Li, J. Sun, and W. Ma, “Joint estimation and prediction of city-wide delivery demand: A large language model empowered graph-based learning approach,”Transportation Research Part E: Logistics and Transportation Review, vol. 197, p. 104075, 2025. 1, 2, 4, 5

work page 2025
[5]

Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

Y . Li, R. Yu, C. Shahabi, and Y . Liu, “Diffusion convolutional re- current neural network: Data-driven traffic forecasting,”arXiv preprint arXiv:1707.01926, 2017. 1, 2, 4, 5

work page internal anchor Pith review Pith/arXiv arXiv 2017
[6]

Graph WaveNet for Deep Spatial-Temporal Graph Modeling

Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph wavenet for deep spatial-temporal graph modeling,”arXiv preprint arXiv:1906.00121,

work page internal anchor Pith review Pith/arXiv arXiv 1906
[7]

Con- necting the dots: Multivariate time series forecasting with graph neural networks,

Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, “Con- necting the dots: Multivariate time series forecasting with graph neural networks,” inProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 753–763. 1, 2, 4, 5

work page 2020
[8]

Graph deep learning for time series forecasting,

A. Cini, I. Marisca, D. Zambon, and C. Alippi, “Graph deep learning for time series forecasting,”arXiv preprint arXiv:2310.15978, 2023. 1

work page arXiv 2023
[9]

Domain adversarial spatial-temporal network: A transferable frame- work for short-term traffic forecasting across cities,

Y . Tang, A. Qu, A. H. Chow, W. H. Lam, S. C. Wong, and W. Ma, “Domain adversarial spatial-temporal network: A transferable frame- work for short-term traffic forecasting across cities,” inProceedings of the 31st ACM international conference on information & knowledge management, 2022, pp. 1905–1915. 1

work page 2022
[10]

Language models represent space and time,

W. Gurnee and M. Tegmark, “Language models represent space and time,”arXiv preprint arXiv:2310.02207, 2023. 1, 2

work page arXiv 2023
[11]

Geollm: Extracting geospatial knowledge from large language models,

R. Manvi, S. Khanna, G. Mai, M. Burke, D. Lobell, and S. Ermon, “Geollm: Extracting geospatial knowledge from large language models,” arXiv preprint arXiv:2310.06213, 2023. 1, 2

work page arXiv 2023
[12]

Geolocation representation from large language models are generic enhancers for spatio-temporal learning,

J. He, T. Nie, and W. Ma, “Geolocation representation from large language models are generic enhancers for spatio-temporal learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 16, 2025, pp. 17 094–17 104. 1, 2

work page 2025
[13]

A poisson-based distribution learning framework for short-term prediction of food de- livery demand ranges,

J. Liang, J. Ke, H. Wang, H. Ye, and J. Tang, “A poisson-based distribution learning framework for short-term prediction of food de- livery demand ranges,”IEEE Transactions on Intelligent Transportation Systems, 2023. 2

work page 2023
[14]

A survey on service route and time prediction in instant delivery: Taxonomy, progress, and prospects,

H. Wen, Y . Lin, L. Wu, X. Mao, T. Cai, Y . Hou, S. Guo, Y . Liang, G. Jin, Y . Zhaoet al., “A survey on service route and time prediction in instant delivery: Taxonomy, progress, and prospects,”IEEE Transactions on Knowledge and Data Engineering, 2024. 2

work page 2024
[15]

On the equivalence between temporal and static equivariant graph representations,

J. Gao and B. Ribeiro, “On the equivalence between temporal and static equivariant graph representations,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 7052–7076. 2, 4, 5

work page 2022
[16]

Inductive graph neural networks for spatiotemporal kriging,

Y . Wu, D. Zhuang, A. Labbe, and L. Sun, “Inductive graph neural networks for spatiotemporal kriging,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 5, 2021, pp. 4478–

work page 2021
[17]

Filling the g_ap_s: Multivari- ate time series imputation by graph neural networks,

A. Cini, I. Marisca, and C. Alippi, “Filling the g_ap_s: Multivari- ate time series imputation by graph neural networks,”arXiv preprint arXiv:2108.00298, 2021. 2, 4, 5

work page arXiv 2021
[18]

Inductive and adaptive graph convolution networks equipped with constraint task for spatial–temporal traffic data kriging,

T. Wei, Y . Lin, S. Guo, Y . Lin, Y . Zhao, X. Jin, Z. Wu, and H. Wan, “Inductive and adaptive graph convolution networks equipped with constraint task for spatial–temporal traffic data kriging,”Knowledge- Based Systems, vol. 284, p. 111325, 2024. 2

work page 2024
[19]

Retrieval- augmented generation for knowledge-intensive nlp tasks,

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020. 2

work page 2020
[20]

Semi-Supervised Classification with Graph Convolutional Networks

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016. 3

work page internal anchor Pith review Pith/arXiv arXiv 2016
[21]

Lade: The first comprehensive last-mile delivery dataset from industry,

L. Wu, H. Wen, H. Hu, X. Mao, Y . Xia, E. Shan, J. Zhen, J. Lou, Y . Liang, L. Yanget al., “Lade: The first comprehensive last-mile delivery dataset from industry,”arXiv preprint arXiv:2306.10675, 2023. 4

work page arXiv 2023
[22]

Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting

B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,”arXiv preprint arXiv:1709.04875, 2017. 4, 5

work page internal anchor Pith review Pith/arXiv arXiv 2017
[23]

Spatial aggregation and temporal convolution networks for real-time kriging,

Y . Wu, D. Zhuang, M. Lei, A. Labbe, and L. Sun, “Spatial aggregation and temporal convolution networks for real-time kriging,”arXiv preprint arXiv:2109.12144, 2021. 5 6

work page arXiv 2021

[1] [1]

Effects of urban delivery restric- tions on traffic movements,

G. Yannis, J. Golias, and C. Antoniou, “Effects of urban delivery restric- tions on traffic movements,”Transportation Planning and Technology, vol. 29, no. 4, pp. 295–311, 2006. 1

work page 2006

[2] [2]

Real-time demand forecast- ing for an urban delivery platform,

A. Hess, S. Spinler, and M. Winkenbach, “Real-time demand forecast- ing for an urban delivery platform,”Transportation Research Part E: Logistics and Transportation Review, vol. 145, p. 102147, 2021. 1, 2

work page 2021

[3] [3]

Autonomous robot- driven deliveries: A review of recent developments and future direc- tions,

S. Srinivas, S. Ramachandiran, and S. Rajendran, “Autonomous robot- driven deliveries: A review of recent developments and future direc- tions,”Transportation research part E: logistics and transportation review, vol. 165, p. 102834, 2022. 1

work page 2022

[4] [4]

Joint estimation and prediction of city-wide delivery demand: A large language model empowered graph-based learning approach,

T. Nie, J. He, Y . Mei, G. Qin, G. Li, J. Sun, and W. Ma, “Joint estimation and prediction of city-wide delivery demand: A large language model empowered graph-based learning approach,”Transportation Research Part E: Logistics and Transportation Review, vol. 197, p. 104075, 2025. 1, 2, 4, 5

work page 2025

[5] [5]

Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting

Y . Li, R. Yu, C. Shahabi, and Y . Liu, “Diffusion convolutional re- current neural network: Data-driven traffic forecasting,”arXiv preprint arXiv:1707.01926, 2017. 1, 2, 4, 5

work page internal anchor Pith review Pith/arXiv arXiv 2017

[6] [6]

Graph WaveNet for Deep Spatial-Temporal Graph Modeling

Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph wavenet for deep spatial-temporal graph modeling,”arXiv preprint arXiv:1906.00121,

work page internal anchor Pith review Pith/arXiv arXiv 1906

[7] [7]

Con- necting the dots: Multivariate time series forecasting with graph neural networks,

Z. Wu, S. Pan, G. Long, J. Jiang, X. Chang, and C. Zhang, “Con- necting the dots: Multivariate time series forecasting with graph neural networks,” inProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 753–763. 1, 2, 4, 5

work page 2020

[8] [8]

Graph deep learning for time series forecasting,

A. Cini, I. Marisca, D. Zambon, and C. Alippi, “Graph deep learning for time series forecasting,”arXiv preprint arXiv:2310.15978, 2023. 1

work page arXiv 2023

[9] [9]

Domain adversarial spatial-temporal network: A transferable frame- work for short-term traffic forecasting across cities,

Y . Tang, A. Qu, A. H. Chow, W. H. Lam, S. C. Wong, and W. Ma, “Domain adversarial spatial-temporal network: A transferable frame- work for short-term traffic forecasting across cities,” inProceedings of the 31st ACM international conference on information & knowledge management, 2022, pp. 1905–1915. 1

work page 2022

[10] [10]

Language models represent space and time,

W. Gurnee and M. Tegmark, “Language models represent space and time,”arXiv preprint arXiv:2310.02207, 2023. 1, 2

work page arXiv 2023

[11] [11]

Geollm: Extracting geospatial knowledge from large language models,

R. Manvi, S. Khanna, G. Mai, M. Burke, D. Lobell, and S. Ermon, “Geollm: Extracting geospatial knowledge from large language models,” arXiv preprint arXiv:2310.06213, 2023. 1, 2

work page arXiv 2023

[12] [12]

Geolocation representation from large language models are generic enhancers for spatio-temporal learning,

J. He, T. Nie, and W. Ma, “Geolocation representation from large language models are generic enhancers for spatio-temporal learning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 16, 2025, pp. 17 094–17 104. 1, 2

work page 2025

[13] [13]

A poisson-based distribution learning framework for short-term prediction of food de- livery demand ranges,

J. Liang, J. Ke, H. Wang, H. Ye, and J. Tang, “A poisson-based distribution learning framework for short-term prediction of food de- livery demand ranges,”IEEE Transactions on Intelligent Transportation Systems, 2023. 2

work page 2023

[14] [14]

A survey on service route and time prediction in instant delivery: Taxonomy, progress, and prospects,

H. Wen, Y . Lin, L. Wu, X. Mao, T. Cai, Y . Hou, S. Guo, Y . Liang, G. Jin, Y . Zhaoet al., “A survey on service route and time prediction in instant delivery: Taxonomy, progress, and prospects,”IEEE Transactions on Knowledge and Data Engineering, 2024. 2

work page 2024

[15] [15]

On the equivalence between temporal and static equivariant graph representations,

J. Gao and B. Ribeiro, “On the equivalence between temporal and static equivariant graph representations,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 7052–7076. 2, 4, 5

work page 2022

[16] [16]

Inductive graph neural networks for spatiotemporal kriging,

Y . Wu, D. Zhuang, A. Labbe, and L. Sun, “Inductive graph neural networks for spatiotemporal kriging,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 5, 2021, pp. 4478–

work page 2021

[17] [17]

Filling the g_ap_s: Multivari- ate time series imputation by graph neural networks,

A. Cini, I. Marisca, and C. Alippi, “Filling the g_ap_s: Multivari- ate time series imputation by graph neural networks,”arXiv preprint arXiv:2108.00298, 2021. 2, 4, 5

work page arXiv 2021

[18] [18]

Inductive and adaptive graph convolution networks equipped with constraint task for spatial–temporal traffic data kriging,

T. Wei, Y . Lin, S. Guo, Y . Lin, Y . Zhao, X. Jin, Z. Wu, and H. Wan, “Inductive and adaptive graph convolution networks equipped with constraint task for spatial–temporal traffic data kriging,”Knowledge- Based Systems, vol. 284, p. 111325, 2024. 2

work page 2024

[19] [19]

Retrieval- augmented generation for knowledge-intensive nlp tasks,

P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschelet al., “Retrieval- augmented generation for knowledge-intensive nlp tasks,”Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020. 2

work page 2020

[20] [20]

Semi-Supervised Classification with Graph Convolutional Networks

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,”arXiv preprint arXiv:1609.02907, 2016. 3

work page internal anchor Pith review Pith/arXiv arXiv 2016

[21] [21]

Lade: The first comprehensive last-mile delivery dataset from industry,

L. Wu, H. Wen, H. Hu, X. Mao, Y . Xia, E. Shan, J. Zhen, J. Lou, Y . Liang, L. Yanget al., “Lade: The first comprehensive last-mile delivery dataset from industry,”arXiv preprint arXiv:2306.10675, 2023. 4

work page arXiv 2023

[22] [22]

Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting

B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,”arXiv preprint arXiv:1709.04875, 2017. 4, 5

work page internal anchor Pith review Pith/arXiv arXiv 2017

[23] [23]

Spatial aggregation and temporal convolution networks for real-time kriging,

Y . Wu, D. Zhuang, M. Lei, A. Labbe, and L. Sun, “Spatial aggregation and temporal convolution networks for real-time kriging,”arXiv preprint arXiv:2109.12144, 2021. 5 6

work page arXiv 2021