TGDT: A Temporal Graph-based Digital Twin for Urban Traffic Corridors

Jeremy Dilmore; Nooshin Yousefzadeh; Rahul Sengupta; Sanjay Ranka

REVIEW 5 major objections 3 minor 23 references

TGDT: A Temporal Graph-based Digital Twin for Urban Traffic Corridors

T0 review · 5 major / 3 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read TGDT is a temporal graph-based digital twin that simultaneously estimates corridor travel time, intersection-level queue length and waiting time, and intervening traffic volumes from minimal inputs, reporting travel-time error around 24…

desk verdict Modular GAT/TCN corridor surrogate is a practical engineering integration, but the paper's headline error claims are contradicted by its own Table 1. read the letter →

arxiv 2504.18008 v2 pith:6SDD4FDI submitted 2025-04-25 cs.LG

Nooshin Yousefzadeh , Rahul Sengupta , Jeremy Dilmore , Sanjay Ranka This is my paper · ORCID

classification cs.LG

keywords digitaltwinurbantrafficcorridorgraphattentionnetworktemporalconvolutionaltraveltimeestimationqueuelengthwaitingsignaloptimization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TGDT is a deep-learning digital twin for signalized urban corridors that estimates several traffic measures at once: bidirectional corridor travel time, per-phase maximum queue length and waiting time at each intersection, and intervening traffic volumes. The paper claims that a modular network combining graph attention layers with temporal convolutions, trained only on interval-aggregated inflows, signal timing plans, and driving behavior parameters, achieves travel-time errors around 24 seconds, waiting-time errors around 100 seconds, and maximum queue-length and volume errors of 4 and 1.5 vehicles on 5-minute intervals. Because the model runs fully parallelized and evaluates over a thousand scenarios in under a minute, the authors argue it can replace slow micro-simulation for real-time signal optimization and corridor assessment. Validation is done on 50,000 hours of synthetic logs from a calibrated micro-simulator of an eight-intersection arterial corridor, benchmarked against earlier graph-based and multi-task digital twins.

What carries the argument

The load-bearing mechanism is a two-level graph representation of the corridor: nodes are intersections, edges are directional road segments, and each scenario becomes a directed acyclic graph with node features holding 5-minute inflow volumes and edge features holding distances, turning counts, densities, and driving-behavior parameters. A static graph feeds an inflow-imputation module built on self-attention and graph attention layers; a dynamic graph, whose edge features evolve over ten time steps, feeds the travel-time module; the learned spatiotemporal representation is then upsampled by transposed convolutions and passed through a CNN encoder to produce queue-length and waiting-time series. Intermediate fusion merges node and edge embeddings, and three separate optimizers update the modules sequentially, which the paper says avoids gradient interference and makes the architecture extensible to more intersections and measures.

What would settle it

Run TGDT on the same corridor with ground-truth queue lengths, waiting times, and travel times collected from video detection, in-road sensors, or probe vehicles over multiple days, and check whether the 5-minute interval estimates stay within the reported bounds (travel time within 24 seconds, waiting time within 100 seconds, queue length within 4 vehicles, volume within 1.5 vehicles); if the real-world errors exceed these bounds systematically, the claim of serving as a real-time digital twin is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that corridor-level and intersection-level measures of effectiveness can be estimated concurrently by one architecture: a graph attention network captures spatial dependencies among intersections and directional road segments, temporal convolutional layers turn those representations into per-phase time series, and a sequential optimization scheme trains each module (inflow imputation, travel time, queue length, waiting time) in a fixed order so that higher-level estimates inform lower-level ones. On an eight-intersection arterial corridor, TGDT reports mean absolute percentage errors around 2.5% for travel time, a Hellinger distance of 0.08 and a normalized Earth mover's distance of 0.04 for travel-time distributions, waiting-time MAPE near 5.5%, and maximum errors of 4 and 1.5 vehicles for queue length and directional volume per 5-minute interval. The paper positions this as an interpretable, scalable surrogate for micro-simulation that needs only commonly collected inputs.

Load-bearing premise

The central claim rests on the assumption that the calibrated micro-simulator reproduces real traffic dynamics closely enough that a model trained on its logs estimates real-world queue lengths, waiting times, and travel times accurately; the paper does not compare predictions against real-world measurements.

Editorial extensions

If this is right

If the reported accuracy holds, traffic agencies can screen thousands of signal-timing plans in seconds, making real-time adaptive signal control feasible without running micro-simulations.
The modular sequential design means adding a new measure of effectiveness or a new intersection should only require retraining or fine-tuning the corresponding module, not the whole network.
The minimal input set of interval inflows, signal plans, turning ratios, distances, and behavior parameters lowers the data barrier for corridors that lack dense sensing.
Corridor-level travel time and intersection-level queue and waiting estimates are produced in lockstep, so a single model can support both route-level and signal-level decisions.
Because it runs in about a minute for 1,000 scenarios, TGDT can serve as the inner-loop surrogate for optimization algorithms that search over signal offsets, cycle lengths, and green splits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the simulator calibration faithfully reproduces the real corridor, the same training pipeline could be retargeted to a new corridor by re-running simulation and retraining; the paper does not demonstrate zero-shot transfer to unseen corridors.
The sequential optimization imposes an implicit causal chain from inflow to travel time to queue and waiting estimates, which could be exploited for error diagnosis: a failure at the queue-length stage can be traced back to the travel-time or inflow module.
A natural next test is to compare TGDT's outputs against real-world detector and probe data on the same corridor; if the sim-to-real gap is small, the framework becomes a practical optimization engine, and if not, the reported error bounds are only valid in the simulated world.
The same architecture could be extended to predict derived measures like delay, emissions, or fuel consumption by adding output heads, since those quantities are correlated with the already-estimated queue and travel-time series.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Two things you should know about arXiv:2504.18008. First, the core contribution is a modular deep-learning surrogate: a GAT-based imputation module feeds a GAT-based travel-time regressor, and temporal CNNs produce per-phase queue length and waiting time outputs, trained sequentially. That integration appears new and is practically sensible. Second, the paper's headline accuracy claims do not hold up against its own Table 1. The abstract and conclusion say TGDT achieves a maximum error of 4 and 1.5 vehicles for queue length and volume at every 5-minute interval. Table 1 reports MAE of 21.152 vehicles for maximum queue length and 5.2843 vehicles for intervening volume under the same 5-minute setting. No definition, footnote, or test-set note reconciles those numbers. This is the first thing to fix before any other evaluation discussion.

What the paper does well: the architecture is cleanly decomposed, the sequential optimization is a pragmatic answer to gradient interference in multi-output models, and the authors generated a large SUMO dataset (50,000 hours) with OD matrices inferred from real detectors and GPS. They also report distributional metrics (Hellinger, EMD) and show robustness under cycle-length, volume, and green-split variations. The GitHub link means the code is inspectable. That counts.

Soft spots beyond the contradiction: the numerical baselines are all the authors' own prior models (MTDT, GAT-AE, FDGNN), so "outperforms state-of-the-art" is really "outperforms our previous versions." There are no error bars or repeated runs. All evaluation is synthetic; no real-world MOE measurements are used, so transferability remains an assumption. Also minor: the intro says 9 intersections, elsewhere 8; the corridor is both 8 and 10 miles in different places. These are sloppy details but not fatal.

Who this is for: traffic engineers and applied ML people who want a fast surrogate for corridor signal optimization. The architecture is worth reading, and the code could be reused. As submitted, the evaluation is not trustworthy as reported. A serious referee should see it, because the problem is important and the method is concrete, but the authors must resolve the accuracy-claim discrepancy, add independent baselines or clearly label the comparison, and ideally include at least one real-world validity check before acceptance.

Recommendation: send to peer review with a major-revision decision.

Referee Report

5 major / 3 minor

Summary. The paper proposes TGDT, a modular deep learning digital twin for urban traffic corridors, combining graph attention networks, temporal convolutions, and transposed convolutions to estimate corridor-level travel time and intersection/phase-level maximum queue length and waiting time. The model is trained and evaluated on 50,000 hours of SUMO-simulated traffic for a Florida SR 436 corridor, using inferred OD matrices and real signal/behavior parameters. The authors report low errors on simulated test data, robustness under varying cycle lengths, traffic volumes, and green-time allocations, and a scalability claim of evaluating 1,000 scenarios in under a minute. The paper also describes sequential optimization over modular losses and provides a public code link.

Significance. If the reported accuracy and scalability hold, a modular GNN/TCN surrogate for corridor-level MOE estimation would be practically useful for real-time traffic signal optimization. The use of a large SUMO dataset, a direction-aware graph representation, and a modular architecture with sequential optimization are reasonable design choices, and the code release is a strength. However, the paper's headline accuracy claims are contradicted by its own Table 1, and the evaluation lacks real-world MOE ground truth, independent baselines, and repeated-run statistics. As presented, the central quantitative claims are not established, although the framework itself may be salvageable with additional analysis.

major comments (5)

[Abstract and Section VI] The headline accuracy claims are contradicted by Table 1. The Abstract and Section VI state maximum errors of 4 vehicles for queue length and 1.5 vehicles for intervening volume at every 5-minute interval, but Table 1's total rows for TGDT (w = 5 min) report MAE = 21.152 for Maximum Queue Length and MAE = 5.2843 for Intervening Traffic Volume. No footnote or test-set definition maps the table values to the claimed 4 and 1.5 vehicle errors, so the paper's most prominent quantitative claim is unsupported by its own evidence.
[Section III-B and Section IV] The evaluation uses only synthetic labels generated by SUMO; no real-world MOE measurements are used as ground truth. Since the digital twin is intended for a real corridor, the transferability claim requires at least a discussion of the simulation-to-reality gap or a validation against field data. Without it, the reported errors and the deployment claims in Section VI remain untested.
[Table 1, Intervening Traffic Volume] The Abstract claims TGDT 'outperforms state-of-the-art baselines,' but in Table 1 the GAT-AE baseline achieves MAE = 1.1815 for Intervening Traffic Volume while TGDT achieves MAE = 5.2843, so for this MOE the proposed model is substantially worse. Moreover, all numerical baselines are the authors' own prior models, and no independent baselines or error bars from repeated runs are provided, which weakens the claimed state-of-the-art comparison.
[Section I vs. Section III-C and Section VI] The corridor description is inconsistent: the Introduction says a 9-intersection, 10-mile corridor, while Section III-C specifies a graph with |V| = 8 and |E| = 16 and Section VI says the target arterial is 8 miles long with 8 signalized intersections. These numbers must be reconciled because the graph size and corridor length directly affect the scalability and error claims.
[Section IV, after Table 1] The text says MAPE shows 'approximately 30 seconds for travel time and 100 seconds for waiting time,' but MAPE values in Table 1 are percentages (0.0248 and 0.0551). The absolute error figures appear to be MAE, not MAPE, and no corresponding MAE column is provided for travel and waiting time. Clarify which metric is being reported and add the missing table entries.

minor comments (3)

[Table 1, Maximum Queue Length, TGDT-Short, Cycle Length High] The reported RMSE of 1739.7 equals the reported MSE of 1739.7 in the same row; since RMSE should be the square root of MSE, this appears to be a typo.
[Section II and Figure 2 caption] The text describes four modules (Minf, Mtt, Mql, Mwt), but the Figure 2 caption says the framework consists of three main modules; this should be corrected for consistency.
[References [7] and [8]] References [7] and [8] share the same arXiv identifier and appear to be duplicate listings; the intended GCRNN citation should be distinguished from DCRNN.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TGDT is a supervised surrogate whose outputs are regressed against independent SUMO-simulated MOE targets; the headline error claims conflict with Table 1, but that inconsistency is a correctness issue, not circular reasoning.

full rationale

TGDT is a supervised multi-output regression model trained on 50,000 hours of SUMO logs and evaluated on a held-out split of the same simulated distribution (Sections III and IV, Table 1). The derivation chain is: raw SUMO trajectories are aggregated into graph-structured inputs (Section III-C); the inflow module imputes missing intervening volumes; the travel-time module regresses bidirectional travel time from graph embeddings; the queue-length and waiting-time modules reuse those embeddings to regress intersection-level MOEs (Section II). The targets are produced by the simulator's FCD and log extraction, not by the model's own equations, so no output is definitionally equal to an input or to another output. The sequential-optimization scheme shares learned representations between modules, but each module's loss is computed against independent simulation targets, so this is feature sharing rather than circular definition. The self-citations that appear (Minf extends [20], and the baselines [19], [20], [21]) are direct architectural provenance and direct empirical comparisons; no argument reduces to an unverified uniqueness theorem or to a prior paper's assertion as the sole support for the central claim. The abstract and conclusion claim a "maximum error of 4 and 1.5 vehicles" for queue length and volume, whereas Table 1 reports total MAEs of 21.152 vehicles and 5.2843 vehicles with no stated mapping to the claimed figures; this is an internal-consistency and correctness problem, not circularity. Likewise, training and testing entirely within SUMO raises external-validity concerns for real-world deployment, but that is a transferability issue rather than a circular-derivation issue. Overall, no step in TGDT's claimed estimation chain reduces by construction to its inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the fidelity of SUMO simulation, the adequacy of an 8-node acyclic graph for corridor modeling, the coverage of the simulated scenario distribution, and a set of architecture hyperparameters chosen without ablation. None of these are validated against external real-world MOE data. Learned network weights are the standard fitted objects of supervised regression; they are not presented as physically derived constants.

free parameters (6)

GAT attention head counts (4 and 1) = 4 and 1
Chosen architecture hyperparameters with no sensitivity analysis in the paper.
Hidden embedding dimension = 64
Used for node embeddings in the travel time module; no ablation is reported.
Temporal context length = 10 time steps
Dynamic graph node tensor captures 10 time steps; no ablation is reported.
Aggregation interval w = 5 minutes (variant: 1 minute)
Core temporal resolution of inputs and outputs; results change substantially between w=5 and w=1.
Kernel sizes, output channels, pooling in temporal modules = not reported
Queue length and waiting time modules use Conv1d, MaxPool1d, and ConvTranspose1d layers whose exact sizes are not specified.
Per-module learning rates for sequential optimization = not reported
Three Adam optimizers with distinct learning rates are used; the values are omitted.

assumptions (5)

domain assumption SUMO micro-simulator, calibrated with real signal timings and driving behavior data, faithfully represents traffic dynamics of SR 436 (Section III-A).
All training and test labels come from SUMO; no real-world MOE ground truth is used for validation.
ad hoc to paper An acyclic directed graph with 8 nodes and 16 edges is sufficient to model dependencies for corridor-level MOE estimation (Section III-C).
The graph is fixed to the single study corridor; generalizability to arbitrary intersection counts is asserted, not tested.
domain assumption The 50,000 hours of simulated scenarios, including randomized OD matrices, cover the operational envelope and extreme scenarios relevant to the claim (Section III-A).
No coverage analysis or comparison to observed extreme events is provided.
domain assumption Only vehicles completing the full corridor route contribute to travel time labels (Section III-B).
Discarding partial trips may bias travel time estimates under congested or spillback conditions.
standard math GAT and temporal CNN layers can represent the mapping from the limited feature set to the MOE outputs (Section II).
Standard supervised-learning assumption that the chosen neural architectures can approximate the target function.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TGDT: A Temporal Graph-based Digital Twin for Urban Traffic Corridors." pith.science (2026). https://pith.science/paper/6SDD4FDI

@misc{pith2026250418008,
  author       = {Pith},
  title        = {Pith review of: TGDT: A Temporal Graph-based Digital Twin for Urban Traffic Corridors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6SDD4FDI}},
  note         = {Machine review of arXiv:2504.18008}
}

read the original abstract

Urban congestion at signalized intersections leads to significant delays, economic losses, and increased emissions. Existing deep learning models often lack spatial generalizability, rely on complex architectures, and struggle with real-time deployment. To address these limitations, we propose the Temporal Graph-based Digital Twin (TGDT), a scalable framework that integrates Temporal Convolutional Networks and Attentional Graph Neural Networks for dynamic, direction-aware traffic modeling and assessment at urban corridors. TGDT estimates key Measures of Effectiveness (MOEs) for traffic flow optimization at both the intersection level (e.g., queue length, waiting time) and the corridor level (e.g., traffic volume, travel time). Its modular architecture and sequential optimization scheme enable easy extension to any number of intersections and MOEs. The model outperforms state-of-the-art baselines by accurately producing high-dimensional, concurrent multi-output estimates. It also demonstrates high robustness and accuracy across diverse traffic conditions, including extreme scenarios, while relying on only a minimal set of traffic features. Fully parallelized, TGDT can simulate over a thousand scenarios within a matter of seconds, offering a cost-effective, interpretable, and real-time solution for urban traffic management and optimization.

Figures

Figures reproduced from arXiv: 2504.18008 by the authors.

**Figure 1.** The inputs and outputs of the proposed urban corridor digital twin. TGDT takes input parameters (highlighted in orange), including ingress aggregated traffic waveforms, signal timing parameters (e.g., cycle length, offset, and maximum green duration for each phase), driving behavior parameters (e.g., speed, acceleration, space cushion, lane-changing behavior), turning movement ratios, and the distances between inter… view at source ↗

**Figure 2.** Overview of TGDT framework. This diagram illustrates the architecture of our proposed Digital Twin for urban corridors, which consists of three main modules. Simulation records, extracted from the logs of a microscopic traffic simulator, are transformed into graph-structured data that uniquely represent the corridor’s traffic state for each scenario. The inflow module (Minf) performs a graph imputation task to recon… view at source ↗

**Figure 3.** Visualization outputs of TGDT for a randomly selected traffic scenario. Comparison of actual (red) and predicted (green) curves for several Measures of Effectiveness (MOEs). TGDT takes ingress (inflow) traffic volumes and various parameters, such as traffic signal plans and driving behaviors, to accurately simulate bidirectional travel times throughout urban traffic corridors. It also estimates multi-directional max… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Predicted versus actual travel time distributions. This figure shows the Kernel Density Estimates (left column) and the joint distribution of travel time characteristics (right column) across the target corridor at a certain time step T = 15 min. The results suggest th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages

[1]

J. L. Elman. Finding structure in time. Cognitive science , 14(2):179– 211, 1990

work page 1990
[2]

Z. Fang, Q. Long, G. Song, and K. Xie. Spatial-temporal graph ode networks for traffic flow forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 364– 373, 2021

work page 2021
[3]

Hochreiter and J

S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997

1997
[4]

T. A. T. Institute. 2021 urban mobility report. Technical report, Texas A&M University, 2021. Accessed: 2025-04-13

work page 2021
[5]

D. R. Jones, M. Schonlau, and W. J. Welch. Efficient global optimization of expensive black-box functions. Journal of Global optimization , 13:455–492, 1998

work page 1998
[6]

LeCun, L

Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278– 2324, 1998

1998
[8]

Y . Li, R. Yu, C. Shahabi, and Y . Liu. Graph convolutional recur- rent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 7(8), 2017

arXiv 2017
[9]

Z. Lin, M. Li, Z. Zheng, Y . Cheng, and C. Yuan. Self-attention convlstm for spatiotemporal prediction. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 11531–11538, 2020

work page 2020

Show all 23 references

[10]

P. A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Fl ¨otter¨od, R. Hilbrich, L. L ¨ucken, J. Rummel, P. Wagner, and E. Wiessner. Microscopic Traffic Simulation Using SUMO. In IEEE Intelligent Transportation Systems Conference (ITSC) . IEEE, 2018

2018
[11]

X. Ma, Z. Tao, Y . Wang, H. Yu, and Y . Wang. Long short-term memory neural network for traffic speed prediction using remote microwave sensor data. Transportation Research Part C: Emerging Technologies , 54:187–197, 2015

2015
[12]

Rossi, B

E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. Bron- stein. Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 , 2020

2006 arXiv
[13]

Siami-Namini, N

S. Siami-Namini, N. Tavakoli, and A. S. Namin. The performance of lstm and bilstm in forecasting time series. In 2019 IEEE International conference on big data (Big Data) , pages 3285–3292. IEEE, 2019

2019
[14]

World urbanization prospects: The 2018 revision, 2018

United Nations, Department of Economic and Social Affairs. World urbanization prospects: The 2018 revision, 2018. Accessed: 2025-04- 13

2018
[15]

Environmental Protection Agency

U.S. Environmental Protection Agency. Fast facts on transportation greenhouse gas emissions, 2022. Accessed: 2025-04-13

2022
[16]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

2017
[17]

H. Wang, R. Zhang, X. Cheng, and L. Yang. Hierarchical traffic flow prediction based on spatial-temporal graph convolutional network. IEEE Transactions on Intelligent Transportation Systems, 23(9):16137–16147, 2022

2022
[18]

Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang. Graph wavenet for deep spatial-temporal graph modeling. arXiv preprint arXiv:1906.00121, 2019

1906 arXiv
[19]

Yousefzadeh, R

N. Yousefzadeh, R. Sengupta, Y . Karnati, A. Rangarajan, and S. Ranka. Mtdt: A multi-task deep learning digital twin. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) , pages 2058–2065, 2024

2024
[20]

Yousefzadeh, R

N. Yousefzadeh, R. Sengupta, Y . Karnati, A. Rangarajan, and S. Ranka. Graph attention network for lane-wise and topology-invariant intersec- tion traffic simulation. IEEE Transactions on Intelligent Transportation Systems, 26(4):5082–5093, 2025

2025
[21]

Yousefzadeh, R

N. Yousefzadeh, R. Sengupta, and S. Ranka. Dynamic graph attention networks for travel time distribution prediction in urban arterial roads. arXiv preprint arXiv:2412.11095 , 2024

2024 arXiv
[22]

Yousefzadeh, M

N. Yousefzadeh, M. T. Thai, and S. Ranka. A comprehensive survey on multilayered graph embedding. TechRxiv Preprint, 2023. Preprint

2023
[23]

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

2017 arXiv
[24]

Zhang, Y

J. Zhang, Y . Zheng, and D. Qi. Deep spatio-temporal residual networks for citywide crowd flows prediction. In Proceedings of the AAAI conference on artificial intelligence , volume 31, 2017

2017

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

J. L. Elman. Finding structure in time. Cognitive science , 14(2):179– 211, 1990

work page 1990

[2] [2]

Z. Fang, Q. Long, G. Song, and K. Xie. Spatial-temporal graph ode networks for traffic flow forecasting. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 364– 373, 2021

work page 2021

[3] [3]

Hochreiter and J

S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997

1997

[4] [4]

T. A. T. Institute. 2021 urban mobility report. Technical report, Texas A&M University, 2021. Accessed: 2025-04-13

work page 2021

[5] [5]

D. R. Jones, M. Schonlau, and W. J. Welch. Efficient global optimization of expensive black-box functions. Journal of Global optimization , 13:455–492, 1998

work page 1998

[6] [6]

LeCun, L

Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278– 2324, 1998

1998

[7] [8]

Y . Li, R. Yu, C. Shahabi, and Y . Liu. Graph convolutional recur- rent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 7(8), 2017

arXiv 2017

[8] [9]

Z. Lin, M. Li, Z. Zheng, Y . Cheng, and C. Yuan. Self-attention convlstm for spatiotemporal prediction. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 11531–11538, 2020

work page 2020