Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Raw OD counts mix demand scale with destination choice; the transferable object is the map from spatial exposures to row-centered choice potentials.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Robust OD flow prediction comes from learning row-centered exposure-to-choice potentials and reconstructing counts as scale times allocation, not from raw-count supervision.

T0 review reviewed 2026-07-12 challenge →

load-bearing objection Solid target-level reframing of OD prediction: scale vs allocation is the real contribution; theory is clean, experiments are useful but not decisive on residual drift. the 3 major comments →

arxiv 2607.03200 v1 pith:VA5GO3DA submitted 2026-07-03 cs.LG stat.ML

OpFlow: Learning Opportunity-Conditioned Choice Potentials for Robust OD Flow Prediction

Changjian Liu , Yong Gao , Yuqing Wang , Leyi Su , Honglei Guo , Zhiyang Wang , Xiaoyu Wang , Fan Zhang This is my paper
classification cs.LG stat.ML
keywords OD flow predictionchoice potentialsspatial interactiondistribution shiftorigin-destinationallocation mechanismopportunity exposurerobust prediction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Deep models trained on raw origin-destination counts often break when cities, policies, or opportunity landscapes change, because those counts mix how much demand an origin produces with where that demand goes. This paper argues that the stable learning target is the exposure-to-choice law: a map from spatial conditions such as opportunities, travel impedance, intervening alternatives, and local competition into row-centered choice potentials whose softmax yields destination allocation. Origin production is predicted by a separate scale branch and flows are reconstructed as scale times allocation. Classical gravity, intervening-opportunity, and radiation laws appear as restricted log-potential cases of the same intensity model, and an OOD bound separates transferable allocation error from residual drift. Controlled synthetic shifts and U.S. county commuting experiments show the approach improves robustness when environments change.

Core claim

Under distribution shift, raw-count supervision cannot distinguish transferable destination-choice mechanisms from environment-specific scale shortcuts. The identifiable transferable target is the row-centered log-potential of an opportunity-conditioned intensity process: spatial exposures map through a shared mechanism to relative destination scores, allocation is the row softmax, and expected OD flows equal origin scale times that allocation. Classical spatial-interaction laws are restricted special cases of this intensity, and OOD allocation error decomposes into training allocation risk (coverage-weighted) plus residual structural drift.

What carries the argument

Opportunity-conditioned choice intensity (Theorem 1) and its identifiable row-centered log-potential: destinations compete via rates that factorize deterrence, opportunity benefit, intervening survival, and local context; OpFlow neuralizes those operators, centers scores by origin, and reconstructs flows with a gradient-isolated scale branch.

Load-bearing premise

After exposures are built, the true map from those exposures to relative destination preferences is shared across environments, with only small leftover differences; if leftovers are large or exposures miss the real drivers of choice, transfer fails.

What would settle it

Train OpFlow and a strong raw-count or generic allocation baseline on shared environments, then raise opportunity-redistribution or opportunity-accessibility correlation-reversal severity; if OpFlow’s row-wise allocation error does not remain clearly lower than the baseline’s as severity rises, the claim that the exposure-to-choice map is the transferable object fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Raw-count training will keep failing under origin-scale heterogeneity even with sophisticated architectures, because scale variance hijacks the objective.
  • Exposure-to-potential maps let destination shares adapt when opportunities are redistributed, instead of freezing historical shares.
  • Gravity, intervening-opportunity, and radiation models are recoverable as fixed parametric special cases of the same operator structure.
  • OD OOD error factors into scale error, transferable allocation error, and residual drift, so robustness needs both scale isolation and mechanism structure.
  • Cross-environment model selection should prioritize held-out allocation risk, not count fit alone.

Where Pith is reading between the lines

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

  • The same scale–allocation split may transfer to other directed interaction matrices under domain shift, such as freight, migration, or platform recommendation flows.
  • If residual drift is large because key choice drivers are unobserved, generic domain-generalization penalties alone will not close the gap; richer exposure operators are required.
  • Active land-use or infrastructure interventions could test whether the learned potentials correctly forecast post-change reallocations.
  • Environment-wise allocation risk penalties may extend beyond commuting to multi-city transfer learning more generally.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper argues that raw OD-count supervision confounds origin production scale with destination allocation, so deep models learn environment-specific shortcuts rather than transferable choice mechanisms. It proposes OpFlow: construct exposure states (opportunity, impedance, intervening rank-prefix, local competition), map them through operator-structured neural channels to row-centered choice potentials, induce allocation by softmax, and reconstruct flows by multiplying with a separately calibrated, gradient-isolated origin scale. Theorem 1 gives an exponential-race / opportunity-intensity foundation; classical gravity, intervening-opportunity, competing-destination, and radiation laws appear as restricted log-potential cases; Assumption 1 and Proposition 1 bound OOD allocation error by transferable training risk plus residual drift. Synthetic interventions (scale severity, opportunity redistribution, exposure-correlation reversal) and a U.S. county commuting benchmark with RandSP/GeoSP/OppSP splits report improved Row-KL, CPC, and Log-RMSE over classical, generic ML, DG, and deep/graph baselines.

Significance. If the transferable-object claim holds, the paper supplies a useful target-level reformulation for robust OD prediction: learn the exposure-to-choice map rather than raw counts, with an explicit scale–allocation decomposition and a clean link to classical spatial-interaction theory. Strengths include a coherent micro-to-aggregate derivation, explicit special-case recovery of classical laws, a correct L1 scale–allocation error decomposition, carefully designed synthetic interventions that isolate scale hijacking and correlation shortcuts, and consistent (if modest) gains on a multi-environment real commuting benchmark. The work is relevant to urban analytics, transportation, and OOD spatiotemporal learning, and the mechanism-constrained architecture is a concrete alternative to unrestricted graph/generative OD models.

major comments (3)
  1. Assumption 1 and Proposition 1 are load-bearing for the claim that the exposure-to-choice map is the transferable object. The synthetic DGP is built from the same opportunity-conditioned intensity family as Theorem 1 / Eq. (3) (oracle g with opportunity, impedance, intervening, local terms plus Gaussian Δ), so Q1–Q4 primarily stress coverage and correlation of exposures OpFlow is designed to reconstruct, not large structural misspecification of F*. Real-world evidence is only cross-county commuting with socio-demo/POI/network features; average Row-KL improves from ~0.41 (PAIRMLP) to 0.377. That is consistent with modest residual drift but does not isolate whether Δ is small enough for the bound to explain the gains, nor whether omitted drivers (mode, policy, unobserved quality) make residual drift dominate under stronger shifts. A misspecification or residual-drift diagnostic (e.g., held
  2. Table 2 and the real-world protocol: gains are consistent but modest, and the strongest generic baselines (PAIRMLP / SCALEMLP) already implement allocation or scale awareness. The paper should clarify how much of OpFlow’s advantage is (i) scale isolation with stop-gradient, (ii) structured exposure operators (rank-prefix, local field), versus (iii) the robust multi-environment objective (max + V-REx). Without a real-world ablation parallel to synthetic Q4, it remains unclear which component is load-bearing under GeoSP/OppSP, and whether the mechanism story is necessary for the reported robustness.
  3. Methodology / Scale-Isolated Reconstruction and Environment-Robust Training: free parameters (τ, T, ω, gate scale ρ, loss weights λ_scale/λ_count/λ_op/λ_vrex, LEO/K-fold selection) are numerous. The manuscript should report sensitivity of OOD metrics to these choices and confirm that LEO selection uses only allocation risk (not count metrics) so that model selection does not re-entangle scale. Without this, the robustness claim is harder to separate from careful multi-environment tuning.
minor comments (5)
  1. Notation: environment superscript e is sometimes suppressed and sometimes retained; a short consistency note would help when reading Theorem 1, Assumption 1, and the methodology equations together.
  2. Figure 4 panels are dense; axis labels and severity definitions for Q1–Q3 should be fully self-contained in the caption so the interventions can be read without the main text.
  3. Table 1 and the appendix classical-law derivations are useful; a one-line pointer in the main text that radiation’s M_i is distinct from the learned origin context m_i would avoid confusion.
  4. Related work could more sharply contrast target-level reformulation (what is supervised) with domain-alignment / invariant-representation methods that keep the raw-count target fixed.
  5. Clarify whether candidate sets D_i are fixed by geography or truncated by K-nearest; this affects row-centering and the interpretation of intervening exposure.

Circularity Check

1 steps flagged

No load-bearing circular derivation; only mild experimental self-alignment of the synthetic DGP with OpFlow’s operator family.

specific steps
  1. other [Synthetic Experiments / DGP and interventions; cf. Eq. (3) and Eq. (7)]
    "The oracle log potential is designed according to Eq. (3): g^{e,∗}_{ij} = a_i O^e_j − b_i C_{ij} − β_s S^e_{ij} + β_ℓ L^e_j + Δ^e_{ij}, with standardized components, Gaussian drift Δ^e_i, and context-aware coefficients a_i, b_i."

    OpFlow’s uncentered score (Eq. 7) uses the same additive operator channels (origin-gated benefit, deterrence, intervening survival, local competition). Synthetic OOD wins and PMSE recovery therefore evaluate a well-specified member of the model class under exposure shifts, not an independent first-principles prediction. This is experimental self-alignment, not a definitional collapse of the theory or real-world results.

full rationale

The paper’s core chain is not circular. Eq. (2) is a standard scale–allocation decomposition of OD means; Theorem 1 derives softmax allocation from an exponential-race intensity with sequential filters and is proved in the appendix (including equivalence to Gumbel RUM), not assumed as the conclusion. Classical laws (Table 1) are obtained by restricting D, C, S, B in Eq. (3)—a genuine specialization argument, not reverse-engineering from a fit. Identifiability of the row-centered potential (Eq. 4) follows from softmax shift-invariance, a standard fact. Assumption 1 and Proposition 1 are stated as assumptions/bounds with residual-drift terms, not claimed as forced uniqueness theorems. Training supervises allocation against empirical row shares and scale against origin totals on held-out environments; real-world gains are measured on external cross-county splits. Citations (Stouffer, Fotheringham, Simini, Luce/McFadden, IRM/V-REx, etc.) are external literature, not self-citation uniqueness chains. The only mild circularity-adjacent issue is experimental: the synthetic oracle log-potential is explicitly “designed according to Eq. (3)” with the same additive channels OpFlow parameterizes, so synthetic mechanism recovery and Q1–Q4 stress tests are partly in-model-class rather than severe misspecification tests. That weakens the force of synthetic evidence but does not make the theoretical claims or real-world evaluation reduce to their inputs by construction. Score 1 reflects that minor experimental self-alignment only.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The central claim rests on a domain modeling package (sequential opportunity filters → multiplicative intensity → softmax allocation), the invariance of the exposure-to-choice map after conditioning on engineered exposures, and several training/architecture knobs. Classical SI and discrete-choice results are imported; the new objects are the row-centered potential target, the operator stack, and the scale-isolation training design. Free parameters are loss and operator temperatures/weights rather than a single fitted physical constant.

free parameters (5)
  • Softmax temperature τ
    Controls allocation sharpness; appears in Theorem 1 and the learned softmax; not fixed by theory.
  • Loss weights λ_scale, λ_count, λ_op, λ_vrex
    Trade off allocation risk, scale fit, count calibration, and operator monotonicity / variance penalty in Eq. (10) and (9); chosen for training, not derived.
  • Soft ranking temperature T and local kernel bandwidth ω
    Define intervening-opportunity and local-competition exposures; continuous design choices that shape Z.
  • Label-smoothing α and origin-gate scale ρ
    Smoothing for full-support allocation targets and gate amplitude 1+ρ tanh(·); training hyperparameters.
  • Network widths/depths and LEO/K-fold selection
    Architecture and model-selection protocol affect reported OOD numbers; not uniquely determined by the theory.
axioms (6)
  • domain assumption Assumption 1: ψ^{e,⋆}_i = H_i F*(Z^e_i) + Δ^e_i with shared F* and bounded row-centered residual drift across environments.
    Load-bearing invariance claim for transfer; stated under Invariant Choice Mechanisms and used in Prop. 1.
  • domain assumption Destination choice is an independent exponential race over effective-opportunity arrivals with multiplicative sequential filters (deterrence, survival, benefit, local context).
    Foundation of Theorem 1; standard-ish in race/RUM models but specific factorization is modeling choice.
  • domain assumption Aggregate OD mean factors as origin production scale times allocation probability (Eq. 2).
    Standard multinomial/Poisson trip generation identity; enables scale-allocation reconstruction.
  • standard math Softmax of row-centered log-potentials is the identifiable allocation target (shift invariance of utilities).
    Standard discrete-choice identifiability; motivates H_i centering in Eq. (4).
  • standard math Deployment exposure distribution is absolutely continuous w.r.t. training with density ratio ≤ ρ.
    Used to convert training allocation risk into the transferable term of Prop. 1.
  • ad hoc to paper Physical monotonicity priors: deterrence and intervening exposure should not increase utility; opportunity should not decrease it (soft ReLU penalties).
    Regularizer Ω_op; plausible but not required by the intensity theorem.
invented entities (3)
  • Opportunity-conditioned choice intensity A*_ij and its neural log-potential operators no independent evidence
    purpose: Provide a micro-to-aggregate interface that unifies classical SI laws and neural allocation learning.
    Defined in Theorem 1 and Eq. (3)/(7); classical laws are recovered as restrictions, but the full neural operator factorization is paper-specific.
  • Row-centered choice potential as the primary learning target independent evidence
    purpose: Remove origin-specific nuisance scale and make the transferable allocation object identifiable.
    Eq. (4) and methodology; standard centering idea applied as the supervised mechanism target for OD OOD.
  • Scale-isolated reconstruction with stop-gradient pooling of potentials no independent evidence
    purpose: Prevent scale loss from rewriting allocation potentials and enforce structural decoupling.
    Eq. (8) and L_count with detached π; engineering entity central to claimed robustness.

reviewed 2026-07-12 · how reviews work

0 comments
Cite this review

Pith. "Pith review of OpFlow: Learning Opportunity-Conditioned Choice Potentials for Robust OD Flow Prediction." pith.science (2026). https://pith.science/paper/VA5GO3DA

@misc{pith2026260703200,
  author       = {Pith},
  title        = {Pith review of: OpFlow: Learning Opportunity-Conditioned Choice Potentials for Robust OD Flow Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VA5GO3DA}},
  note         = {Machine review of arXiv:2607.03200}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Origin-destination (OD) flow prediction is central to urban analytics, yet deep models trained on raw counts remain vulnerable to distribution shift. The core problem is that raw count supervision cannot distinguish transferable choice mechanisms from environment-specific shortcuts. Raw OD count mixes two objects: how much demand an origin produces and how that demand is allocated across destinations. We argue that the transferable object is the exposure-to-choice law that maps spatial conditions to relative destination preferences. We propose OpFlow, a mechanism-constrained framework that learns row-centered choice potentials and reconstructs flows by combining the induced allocation with a separately calibrated origin scale. Under distribution shift, spatial exposures and the induced allocations are allowed to vary; what transfers is the conditional map from exposure states to relative choice potentials. Theoretically, we characterize the identifiable row-centered potential and show that classical spatial interaction laws are restricted log-potential cases. Controlled synthetic shifts and a real-world experiment show OpFlow improves robustness under environment shifts.

Figures

Figures reproduced from arXiv: 2607.03200 by Changjian Liu, Fan Zhang, Honglei Guo, Leyi Su, Xiaoyu Wang, Yong Gao, Yuqing Wang, Zhiyang Wang.

Figure 1
Figure 1. Figure 1: From raw count to transferable mechanisms. Raw OD counts make deployable mechanisms and spurious shortcuts indistinguishable ID. OpFlow instead targets the exposure-to-choice mechanism that remains stable under shift. The Opportunity-Conditioned Choice Mechanism We denote the input covariates as Xe ij = (x e,ori i , x e,dest j , x e,pair ij ), comprising origin, destination and pairwise fea￾tures. While th… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed method. structured neural spatial-interaction operators. These oper￾ators strictly follow the opportunity-conditioned intensity law (Theorem 1), ensuring the learned potential remains grounded in transferable behavioral mechanisms. Operator-Structured Choice Potential Given the input Xe , OpFlow first applies an exposure con￾struction operator O to build exposure state Z e ij . O c… view at source ↗
Figure 3
Figure 3. Figure 3: Synthetic interventions used to evaluate target [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Synthetic OOD results across four interventions. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 6 linked inside Pith

  1. [1]

    1989 , publisher=

    Spatial interaction models: formulations and applications , author=. 1989 , publisher=

    1989

  2. [2]

    International Journal of Digital Earth , volume=

    Integrating trajectory data and demographic characteristics: a trajectory semantic model for predicting travel flow and conducting interaction analysis , author=. International Journal of Digital Earth , volume=. 2024 , publisher=

    2024

  3. [3]

    Physics Reports , volume=

    Human mobility: Models and applications , author=. Physics Reports , volume=. 2018 , publisher=

    2018

  4. [4]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    Urban computing: concepts, methodologies, and applications , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2014 , publisher=

    2014

  5. [5]

    ACM Computing Surveys , volume=

    An interdisciplinary survey on origin-destination flows modeling: Theory and techniques , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  6. [6]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

    Origin-destination matrix prediction via graph convolution: a new perspective of passenger demand modeling , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

  7. [7]

    2020 IEEE 36th International conference on data engineering (ICDE) , pages=

    Predicting origin-destination flow via multi-perspective graph convolutional network , author=. 2020 IEEE 36th International conference on data engineering (ICDE) , pages=. 2020 , organization=

    2020

  8. [8]

    Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

    Continuous-time and multi-level graph representation learning for origin-destination demand prediction , author=. Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

  9. [9]

    Nature Machine Intelligence , volume=

    Shortcut learning in deep neural networks , author=. Nature Machine Intelligence , volume=. 2020 , publisher=

    2020

  10. [10]

    arXiv preprint arXiv:1907.02893 , year=

    Invariant risk minimization , author=. arXiv preprint arXiv:1907.02893 , year=

    Pith/arXiv arXiv 1907

  11. [11]

    American sociological review , volume=

    Intervening opportunities: a theory relating mobility and distance , author=. American sociological review , volume=. 1940 , publisher=

    1940

  12. [12]

    Regional studies , volume=

    Towards a cognitive gravity model: The case of consumer spatial behaviour , author=. Regional studies , volume=. 1981 , publisher=

    1981

  13. [13]

    Environment and Planning A: Economy and Space , volume=

    A new set of spatial-interaction models: the theory of competing destinations , author=. Environment and Planning A: Economy and Space , volume=. 1983 , publisher=

    1983

  14. [14]

    Nature , volume=

    A universal model for mobility and migration patterns , author=. Nature , volume=. 2012 , publisher=

    2012

  15. [15]

    Nature communications , volume=

    A deep gravity model for mobility flows generation , author=. Nature communications , volume=. 2021 , publisher=

    2021

  16. [16]

    Journal of Transport Geography , volume=

    Systematic comparison of trip distribution laws and models , author=. Journal of Transport Geography , volume=. 2016 , publisher=

    2016

  17. [17]

    The Thirteenth International Conference on Learning Representations , year=

    A Large-scale Dataset and Benchmark for Commuting Origin-Destination Flow Generation , author=. The Thirteenth International Conference on Learning Representations , year=

  18. [18]

    International conference on machine learning , pages=

    Out-of-distribution generalization via risk extrapolation (rex) , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  19. [19]

    arXiv preprint arXiv:2007.01434 , year=

    In search of lost domain generalization , author=. arXiv preprint arXiv:2007.01434 , year=

    Pith/arXiv arXiv 2007

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Learning invariant graph representations for out-of-distribution generalization , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    arXiv preprint arXiv:2201.12872 , year=

    Discovering invariant rationales for graph neural networks , author=. arXiv preprint arXiv:2201.12872 , year=

    Pith/arXiv arXiv

  22. [22]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Out-of-distribution generalization on graphs: A survey , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  23. [23]

    1959 , publisher=

    Individual choice behavior , author=. 1959 , publisher=

    1959

  24. [24]

    Conditional logit analysis of qualitative choice behavior , author=

  25. [25]

    Journal of the American Institute of planners , volume=

    How accessibility shapes land use , author=. Journal of the American Institute of planners , volume=. 1959 , publisher=

    1959

  26. [26]

    Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems , pages=

    VisitFrequency-Diffusion: Leveraging Recurrent Visits for Long-Term Individual Trajectory Forecasting , author=. Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems , pages=

  27. [27]

    arXiv preprint arXiv:2306.03390 , year=

    Origin-Destination Network Generation via Gravity-Guided GAN , author=. arXiv preprint arXiv:2306.03390 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2306.04873 , year=

    Complexity-Aware Large Scale Origin-Destination Network Generation via Diffusion Model , author=. arXiv preprint arXiv:2306.04873 , year=

    Pith/arXiv arXiv

  29. [29]

    Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence , pages=

    Cross-City Transfer Learning for Deep Spatio-Temporal Prediction , author=. Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence , pages=. 2019 , doi=

    2019

  30. [30]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Selective Cross-City Transfer Learning for Traffic Prediction via Source City Region Re-Weighting , author=. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=. 2022 , publisher=

    2022

  31. [31]

    Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence , pages=

    When Transfer Learning Meets Cross-City Urban Flow Prediction: Spatio-Temporal Adaptation Matters , author=. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence , pages=. 2022 , doi=

    2022

  32. [32]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    GODDAG: Generating Origin-Destination Flow for New Cities via Domain Adversarial Training , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2023 , doi=

    2023

  33. [33]

    arXiv preprint arXiv:2410.04740 , year=

    Evaluating the Generalization Ability of Spatiotemporal Model in Urban Scenario , author=. arXiv preprint arXiv:2410.04740 , year=

    Pith/arXiv arXiv

  34. [34]

    GeoComputational Modelling: Techniques and Applications , editor =

    Neural Spatial Interaction Models , author =. GeoComputational Modelling: Techniques and Applications , editor =. 2001 , doi =

    2001

  35. [35]

    Geographical Analysis , volume =

    A Methodology for Neural Spatial Interaction Modeling , author =. Geographical Analysis , volume =. 2002 , doi =

    2002

  36. [36]

    Proceedings of the 36th IEEE International Conference on Data Engineering , pages =

    Stochastic Origin-Destination Matrix Forecasting Using Dual-Stage Graph Convolutional, Recurrent Neural Networks , author =. Proceedings of the 36th IEEE International Conference on Data Engineering , pages =. 2020 , publisher =

    2020

  37. [37]

    Neural Computing and Applications , volume =

    Deep Learning for Short-Term Origin--Destination Passenger Flow Prediction under Partial Observability in Urban Railway Systems , author =. Neural Computing and Applications , volume =. 2022 , doi =

    2022

  38. [38]

    Expert Systems with Applications , volume =

    Graph Neural Network for Traffic Forecasting: A Survey , author =. Expert Systems with Applications , volume =. 2022 , doi =

    2022

  39. [39]

    1985 , publisher=

    Discrete choice analysis: theory and application to travel demand , author=. 1985 , publisher=

    1985

  40. [40]

    2012 , publisher=

    A universal model of commuting networks , author=. 2012 , publisher=

    2012

  41. [41]

    Journal of Transport geography , volume=

    Accessibility evaluation of land-use and transport strategies: review and research directions , author=. Journal of Transport geography , volume=. 2004 , publisher=

    2004

  42. [42]

    International conference on machine learning , pages=

    Wilds: A benchmark of in-the-wild distribution shifts , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

This paper was first reviewed by grok-4.5 on July 12, 2026.