Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

A spatio-temporal graph neural network estimates daily traffic volume profiles across an entire road network from speed profiles, static road features, and local topology, with no volume data needed at inference time.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 16:26 UTC pith:DLRF2ZYG

load-bearing objection Useful architecture and dataset for volume-from-speed estimation, but the free-flow identifiability gap and lack of external baselines keep the network-wide claim provisional. the 4 major comments →

arxiv 2512.13758 v1 pith:DLRF2ZYG submitted 2025-12-15 cs.LG cs.AI

Network-Wide Traffic Volume Estimation from Speed Profiles using a Spatio-Temporal Graph Neural Network with Directed Spatial Attention

L\'eo Hein (IFPEN) , Giovanni de Nunzio (IFPEN) , Giovanni Chierchia (LIGM) , Aur\'elie Pirayre (IFPEN) , Laurent Najman (LIGM) This is my paper
classification cs.LG cs.AI
keywords traffic volume estimationspatio-temporal graph neural networksdirected graph attentioninductive learningspeed profilesroad network topologyLyon datasetGEH
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.

This paper tries to establish that traffic volume profiles for every road segment in a city can be estimated from data that is much cheaper to obtain than volume sensors: probe-vehicle speed profiles, static road descriptors, and the road network's topology. It proposes HDA-STGNN, a graph neural network trained on a small set of labeled roads that learns a local mapping from a road's K-hop neighborhood to its daily volume profile. Once trained, the model is inductive and can be applied to any road segment in the network, or potentially in other cities, without needing volume data at inference time. On a year of data from Lyon, the model achieves a mean GEH of 6.09 on held-out sensors, improving over ablations that drop speed, static features, neighborhood context, or directed attention. A sympathetic reader would take away that sensor-free network-wide volume estimation is feasible in principle, with the main caveat being roads that are always in free flow.

Core claim

The paper's central claim is that a daily volume profile for any road segment can be predicted as a function of that segment's local topological context, its static attributes, and the speed profile of the segment and its neighbors, with no volume measurements required at inference time. The authors demonstrate this with HDA-STGNN, whose key design is a dual-branch architecture: one branch processes the spatio-temporal speed tensor through factorized 1D convolutions and directed graph attention, the other processes static descriptors through directed graph attention alone; the two embeddings are merged and decoded into volume profiles. Trained on 80% of Lyon's 653 instrumented road segments

What carries the argument

The central mechanism is the Directed Graph Attention (DGAT) layer applied to an oriented dual graph of the road network, where each node is a directed road segment and edges follow allowed maneuvers. DGAT extends standard graph attention by assigning separate learned parameters (projection matrices, attention vectors, and direction embeddings) to incoming, outgoing, and self-loop edges, so the model can distinguish upstream from downstream traffic states when aggregating neighborhood information. This is combined with symmetric 1D temporal convolutions in a factorized spatio-temporal block, and a second DGAT branch on static descriptors, to produce a fully inductive transformation with no g

Load-bearing premise

Speed profiles carry enough information about traffic demand on a road segment to infer its volume, which fails for roads that are almost always in free-flow, where constant speeds reveal nothing about how many vehicles use the road.

What would settle it

Apply the trained model to a set of road segments known to be in free-flow for the entire day (e.g., overnight or on a highway with no congestion) and compare predicted volumes to actual counts: if GEH degrades sharply for exactly these segments, the speed-only premise is the limiting factor.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Volume estimates can be produced for every road segment in a city, including roads with no sensors, using widely available probe speed data and map attributes.
  • The model transfers inductively to unseen roads and, potentially, to entirely different networks without retraining, since it uses only local features.
  • Explicitly modeling edge direction (upstream/downstream) improves volume estimation over undirected graph attention.
  • Static road descriptors (speed limit, lanes, functional class, etc.) are essential for spatial generalization; speed alone is insufficient.
  • The approach estimates typical weekday volume profiles (daily averages) and degrades on raw daily data, pointing to the need for richer temporal features.

Where Pith is reading between the lines

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

  • The same architecture could be tested on a second city to see whether the learned speed-to-volume mapping transfers across networks; if it does, the method becomes a sensor-free volume estimation baseline.
  • Adding flow-conservation constraints at intersections (the paper notes lack of flow conservation) could reduce local inconsistencies and improve network-wide realism.
  • The free-flow ambiguity is fundamental: in regimes where speed is constant, volume is unidentifiable from speed alone, so the method's ceiling is set by how much demand variation is reflected in speed variation.
  • Combining day-specific probe speeds (rather than weekly averages) with calendar/weather features could extend the approach from typical weekday profiles to day-specific estimation.

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

4 major / 4 minor

Summary. The paper proposes HDA-STGNN, an inductive spatio-temporal graph neural network that estimates daily traffic volume profiles for every road segment in an urban network using only speed profiles, static road attributes, and the local topological graph, with no volume data at inference time. The model is a two-branch architecture: a spatio-temporal branch processes speed time series with 1D convolutions and directed graph attention (DGAT), while a spatial branch processes static descriptors with DGAT. The authors construct a custom Lyon dataset (2021) from HERE Maps geospatial data, traffic counts, and GPS traces, and evaluate on 20% held-out sensors with ablations. The reported best model achieves a mean GEH of 6.09 and 46.61% of GEH>5, and a qualitative network-wide visualization is provided. The central claim is that the learned transformation generalizes inductively to unseen road segments and, potentially, to other networks.

Significance. The problem is well-motivated: volume data on every road is seldom available, while speed profiles and static road attributes are broadly accessible. The proposed architecture is a reasonable, non-trivial combination of directed graph attention and temporal convolutions, and the ablation study (Table II) provides useful evidence that each component contributes. The inductive design is a genuine strength, and the paper is careful to exclude validation sensors from training. However, the significance is limited by several evaluation gaps: there are no external baselines, no error bars across the five seeds, no quantitative validation on truly unlabeled subareas, and the authors themselves concede the free-flow identifiability problem. In addition, the reported mean GEH and %GEH>5 exceed the conventional acceptability threshold that the paper itself cites. If the results hold, the method could be a valuable tool for sensor-scarce cities, but the current evidence does not fully support the network-wide claim.

major comments (4)
  1. [§IV.C, last paragraph] The paper concedes a load-bearing identifiability gap: 'Many of these segments are almost always in a fluid regime, and the model may lack the necessary features to infer true demand levels, information not reflected by constant speed profiles and insufficiently compensated by neighborhood context.' In free flow, speed is near the free-flow speed and flat across the day, so the mapping from speed to volume is many-to-one: a 50 km/h road could carry 100 or 1000 veh/h depending on demand. Since free flow is likely the dominant regime for many urban roads, the 'network-wide' claim is not supported. Provide a quantitative breakdown of performance by regime (e.g., free-flow vs. congested segments) and report the fraction of segments in each regime. Without this, the main claim is only established for the subset of roads with informative speed variation.
  2. [§IV.B, Table II] There are no external baselines. Table II reports only self-ablation variants, so the absolute quality of HDA-STGNN is not benchmarked against fundamental-diagram approaches (cited as [2], [3]), earlier probe-speed ML models ([4], [5]), or standard STGNN baselines. The absence of error bars is also a problem: results are averaged over five seeds, but no standard deviations are reported, so we cannot assess the significance of the differences between ablations. Finally, the paper defines a common threshold for GEH (at most 15% of GEH>5 for satisfactory predictions) but does not discuss that its own result of 46.61% GEH>5 is far above that threshold; this must be addressed to support the 'promising performance' claim.
  3. [§IV.A.4 and §II] The temporal granularity of the inputs does not match the stated goal. Speed profiles are provided by HERE Maps as average profiles for each day of the week, and the authors correspondingly average volume profiles by weekday. Thus the model learns to estimate a 'typical weekday' volume profile, not the volume of any specific day. The claim in Section II that the model can be applied 'for any given day' (and the title's 'from Speed Profiles') is therefore misleading. The raw-volume ablation (Ablation 6) still uses the same averaged speed inputs, so it cannot resolve the mismatch. Either use day-specific speed data or explicitly limit the claims to typical-day estimation.
  4. [§IV.C, Fig. 6 and §IV.B] The network-wide estimation is never quantitatively evaluated on unlabeled roads. Figure 6 is only a visual map, and the paper states 'a quantitative evaluation of unlabeled roads is not possible.' The 20% random sensor split is an interpolation-like test: the held-out sensors are embedded in a graph where many of their neighbors are training sensors. This does not simulate a sensor-free subarea or a fully unmonitored network, yet the conclusion claims applicability to 'unseen road sections.' A leave-areas-out protocol (e.g., removing all sensors in a contiguous region) or a transfer experiment to another city is needed to support the inductive network-wide claim.
minor comments (4)
  1. [§IV.B] Typo: 'Cf = 7static descriptors' should read 'Cf = 7 static descriptors.'
  2. [§IV.B] The hardware description states 'NVIDIA GeForce RTX 4070 GPU and an Intel i5-5600KF CPU' — the Intel model number appears to be a typo, as no such CPU exists. Please correct.
  3. [Throughout] The GEH criterion is written inconsistently ('GEH>5' vs. 'GEH > 5'); unify the formatting.
  4. [§IV.A.2 / §VI] The paper does not state whether the custom Lyon dataset or the implementation will be released. For reproducibility, please include a data/code availability statement or a reason for non-release.

Circularity Check

0 steps flagged

No significant circularity: the volume targets are used only as supervised labels, never as inference inputs, and held-out sensors are genuinely excluded from training.

full rationale

The paper's derivation chain is an empirical supervised-learning pipeline: fθ maps (G_NK, F_NK, P_NK,d) to Q̂, with Q used only as the training target in the Huber loss (Section IV.B). No equation defines the target in terms of the input, and no parameter is fitted directly to held-out targets: 20% of sensors are excluded from training and the same learned fθ is then applied to them (Section IV.B). The weekday-averaged targets and speed profiles share a temporal aggregation, which weakens day-specific generalization, but this is a data-construction limitation, not circularity: the target values are not among the inputs at inference. The free-flow remark in Section IV.C — 'Many of these segments are almost always in a fluid regime, and the model may lack the necessary features to infer true demand levels' — identifies an identifiability/support problem, not equivalence-by-construction. The DGAT components are attributed to independent prior work [20], [21]; no load-bearing self-citation or imported uniqueness claim is present. Therefore the derivation is self-contained with respect to circularity, and concerns about free-flow identifiability belong to correctness risk rather than circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The method relies on standard supervised learning; learned weights are trained, not fitted ad hoc. The listed free parameters are manually tuned hyperparameters, none with sensitivity analysis. The axioms are domain assumptions about traffic behavior and data representativeness that go unvalidated.

free parameters (6)
  • Neighborhood depth K = 2
    Chosen as 'balanced trade-off between model complexity, computation time, and spatial scale' (Section IV.A.4); no sensitivity analysis.
  • Hidden dimension C = 64
    Set for all layers; manually tuned, no ablation.
  • 1D-CNN kernel size p = 9
    Manually chosen; no sensitivity analysis.
  • Number of attention heads = 4
    Manual choice; no sensitivity analysis.
  • Huber loss threshold δ = 50
    Chosen without stated justification (Section IV.B).
  • Dropout rates = 0.1–0.6
    Adjusted per layer type; manual tuning.
axioms (5)
  • domain assumption Traffic volume is stable at hourly scale and exogenous flows are negligible when aggregating road links between intersections
    Section IV.A.3: 'aggregate each set of links connecting two intersections into a single unified road segment, assuming traffic volume is stable at an hourly scale and neglecting exogenous flows.'
  • domain assumption The labeled sensor nodes are representative of unlabeled road segments in feature distribution
    Section IV.A.5: 'differences between distributions appear small enough to expect reliable inference' — this is visually assessed, not quantitatively tested.
  • domain assumption Weekday-averaged HERE Maps speed profiles contain sufficient information to predict typical weekday volume profiles
    The entire dataset is built on this (Section IV.A.4); the authors acknowledge free-flow roads lack informative speed variation.
  • ad hoc to paper A 2-hop neighborhood subgraph of the dual graph captures the local traffic dependencies needed
    Chosen as a balanced trade-off (Section IV.A.4) without empirical justification of sufficiency.
  • domain assumption The directed dual graph representation with direction labels adequately encodes traffic regulations and flow
    Section II: graph constructed from HERE Maps traffic rules; no validation of this representation.

pith-pipeline@v1.3.0-alltime-deepseek · 10689 in / 7717 out tokens · 65122 ms · 2026-08-03T16:26:35.300500+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Network-Wide Traffic Volume Estimation from Speed Profiles using a Spatio-Temporal Graph Neural Network with Directed Spatial Attention." pith.science (2026). https://pith.science/paper/DLRF2ZYG

@misc{pith2026251213758,
  author       = {Pith},
  title        = {Pith review of: Network-Wide Traffic Volume Estimation from Speed Profiles using a Spatio-Temporal Graph Neural Network with Directed Spatial Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLRF2ZYG}},
  note         = {Machine review of arXiv:2512.13758}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Existing traffic volume estimation methods typically address either forecasting traffic on sensor-equipped roads or spatially imputing missing volumes using nearby sensors. While forecasting models generally disregard unmonitored roads by design, spatial imputation methods explicitly address network-wide estimation; yet this approach relies on volume data at inference time, limiting its applicability in sensor-scarce cities. Unlike traffic volume data, probe vehicle speeds and static road attributes are more broadly accessible and support full coverage of road segments in most urban networks. In this work, we present the Hybrid Directed-Attention Spatio-Temporal Graph Neural Network (HDA-STGNN), an inductive deep learning framework designed to tackle the network-wide volume estimation problem. Our approach leverages speed profiles, static road attributes, and road network topology to predict daily traffic volume profiles across all road segments in the network. To evaluate the effectiveness of our approach, we perform extensive ablation studies that demonstrate the model's capacity to capture complex spatio-temporal dependencies and highlight the value of topological information for accurate network-wide traffic volume estimation without relying on volume data at inference time.

Figures

Figures reproduced from arXiv: 2512.13758 by Aur\'elie Pirayre (IFPEN), Giovanni Chierchia (LIGM), Giovanni de Nunzio (IFPEN), Laurent Najman (LIGM), L\'eo Hein (IFPEN).

Figure 1
Figure 1. Figure 1: HDA-STGNN framework and architecture. graph, while the spatial branch applies DGAT layers on the static descriptor graph to capture topology-aware and time￾independent spatial dependencies. The complementary em￾beddings obtained from both branches are then concatenated and fed to fully connected layers to predict the daily volume profiles in the subgraph. Finally, we select the estimated volume profile of … view at source ↗
Figure 2
Figure 2. Figure 2: Examples of speed profiles associated with the measured volume [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Input features representativeness of training nodes. From left [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of estimated traffic volume profiles by HDA-STGNN [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Network-wide hourly volume estimation for Tuesdays at 9a.m.. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Capacity-Aware Deep Learning for Generalizable Traffic Volume Estimation Across Links and Cities

    cs.LG 2026-07 conditional novelty 5.0

    Factoring predicted traffic volume into a learned link capacity and a regime-aware utilization ratio improves hourly volume estimates on unseen links and across cities.

  2. Selecting New Measurement Locations to Diversify Traffic-Pattern Coverage: A Real-World Evaluation for Total Traffic Volume Estimation

    cs.NI 2026-05 unverdicted novelty 5.0

    An algorithm selects traffic counter locations to increase observed traffic-pattern diversity; real-world installation of the chosen counters improved volume estimation accuracy.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    Europe-wide high-spatial resolution air pollution models are improved by including traffic flow estimates on all roads,

    Y . Shen, K. D. Hoogh, O. Schmitz, J. Gulliver, D. Vienneau, R. Vermeulen, G. Hoek, and D. Karssenberg, “Europe-wide high-spatial resolution air pollution models are improved by including traffic flow estimates on all roads,” Atmospheric Environment, vol. 335, no. 120719, 2024

    2024

  2. [2]

    Empirical macroscopic fundamental diagrams: New insights from loop detector and floating car data,

    L. Amb ¨uhl, B. Chopard, Y . Nicolet, A. Marconi, G. Fl ¨otter¨od, M. Bierlaire, and K. W. Axhausen, “Empirical macroscopic fundamental diagrams: New insights from loop detector and floating car data,” in TRB 96th Annual Meeting Compendium of Papers. 96th Annual Meet- ing of the Transportation Research Board (TRB 2017) , Transportation Research Board, p. ...

    work page doi:10.3929/ethz-b-000167171 2017

  3. [3]

    Estimating traffic flow rate on freeways from probe vehicle data and fundamental diagram,

    K. Anuar, F. Habtemichael, and M. Cetin, “Estimating traffic flow rate on freeways from probe vehicle data and fundamental diagram,” in Proceedings of the 2015 IEEE 18th International Conference on Intelligent Transportation Systems (ITSC 2015) , pp. –, Sept. 2015, doi: 10.1109/ITSC.2015.468

    doi:10.1109/itsc.2015.468 2015

  4. [4]

    Traffic flow estimation using probe vehicle data,

    O. Gkountouna, D. Pfoser, and A. Z ¨ufle, “Traffic flow estimation using probe vehicle data,” in Proceedings of the 2020 IEEE 7th International Conference on Data Science and Advanced Analytics (DSAA), Sydney, NSW, Australia, Oct. 2020, doi: 10.1109/DSAA49011.2020.00073

    arXiv 2020

  5. [5]

    Multi-models machine learning methods for traffic flow estimation from Floating Car Data,

    Y . Li, Y . Han, S. Chen, and Y . Wang, “Multi-models machine learning methods for traffic flow estimation from Floating Car Data,” Transportation Research Part C: Emerging Technologies, vol. 132, p. 103389, 2021, doi: 10.1016/j .trc.2021.103389

    arXiv 2021

  6. [6]

    Network topological ef- fects on the macroscopic fundamental diagram,

    W. Wong, S. C. Wong, and H. X. Liu, “Network topological ef- fects on the macroscopic fundamental diagram,” Transportmetrica B: Transport Dynamics , vol. 9, no. 1, pp. 376–398, 2021, doi: 10.1080/21680566.2021.1871484

    arXiv 2021

  7. [7]

    DL-Traff: A survey and benchmark of deep learning models for urban traffic prediction,

    R. Jiang, D. Yin, Z. Wang, Y . Wang, J. Deng, H. Liu, Z. Cai, J. Deng, X. Song, and R. Shibasaki, “DL-Traff: A survey and benchmark of deep learning models for urban traffic prediction,” arXiv preprint arXiv:2108.09091 , 2021

    Pith/arXiv arXiv 2021

  8. [8]

    Spatio-temporal graph neural networks for predictive learning in urban computing: A survey,

    G. Jin, Y . Liang, Y . Fang, Z. Shao, J. Huang, J. Zhang, and Y . Zheng, “Spatio-temporal graph neural networks for predictive learning in urban computing: A survey,” arXiv preprint arXiv:2303.14483 , 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Spatio-temporal graph neural networks: A survey,

    Z. A. Sahili and M. Awad, “Spatio-temporal graph neural networks: A survey,” arXiv preprint arXiv:2301.10569 , 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Evaluating the generalization ability of spatiotemporal model in urban scenario,

    Hongjun Wang, Jiyuan Chen, Tong Pan, Zheng Dong, Lingyu Zhang, Renhe Jiang, Xuan Song, “Evaluating the generalization ability of spatiotemporal model in urban scenario,” arXiv preprint arXiv:2410.04740, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS 2017), Long Beach, CA, USA, 2017, pp. 1025–1035

    2017

  12. [12]

    Comparison be- tween inductive and transductive learning in a real citation net- work using graph neural networks,

    G. Lachaud, P. Conde-Cespedes, and M. Trocan, “Comparison be- tween inductive and transductive learning in a real citation net- work using graph neural networks,” in Proceedings of the 2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM 2022) , 2022, pp. 534-540, doi: 10.1109/ASONAM55673.2022.10068589

    arXiv 2022

  13. [13]

    Network-Wide Traffic Flow Estimation Across Multiple Cities with Global Open Multi- Source Data: A Large-Scale Case Study in Europe and North Amer- ica,

    Z. Hu, Z. Zheng, M. Menendez, and W. Ma, “Network-Wide Traffic Flow Estimation Across Multiple Cities with Global Open Multi- Source Data: A Large-Scale Case Study in Europe and North Amer- ica,” arXiv preprint arXiv:2502.03798 , 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Network-wide Freeway Traffic Estimation Using Sparse Sensor Data: A Dirichlet Graph Auto-Encoder Approach,

    Q. Zhou, Y . Zhang, M. A. Makridis, A. Kouvelas, Y . Wang, and S. Hu, “Network-wide Freeway Traffic Estimation Using Sparse Sensor Data: A Dirichlet Graph Auto-Encoder Approach,” arXiv preprint arXiv:2503.15845 , 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Urban Network-Wide Traffic V olume Estimation Under Sparse De- ployment of Detectors,

    J. Xing, R. Liu, Y . Zhang, C. F. Choudhury, X. Fu, and Q. Cheng, “Urban Network-Wide Traffic V olume Estimation Under Sparse De- ployment of Detectors,” Transportmetrica A: Transport Science , vol. 20, no. 3, pp. 2197511, 2024, doi: 10.1080/23249935.2023.2197511

    arXiv 2024

  16. [16]

    Network-Wide Traffic Flow Estimation with Insufficient V olume Detection and Crowdsourcing Data,

    Z. Zhang, M. Li, X. Lin, and Y . Wang, “Network-Wide Traffic Flow Estimation with Insufficient V olume Detection and Crowdsourcing Data,” Transportation Research Part C: Emerging Technologies , vol. 121, pp. 102870, 2020, doi: 10.1016/j .trc.2020.102870

    arXiv 2020

  17. [17]

    Towards Better Traffic V olume Estimation: Jointly Addressing the Underdetermination and Nonequilibrium Problems with Correlation-Adaptive GNNs,

    T. Nie, G. Qin, Y . Wang, and J. Sun, “Towards Better Traffic V olume Estimation: Jointly Addressing the Underdetermination and Nonequilibrium Problems with Correlation-Adaptive GNNs,” arXiv preprint arXiv:2303.05660 , 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” in Proceedings of the 6th International Conference on Learning Representations (ICLR 2018) , Vancouver, Canada, 2018

    2018

  19. [19]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proceedings of the 5th International Conference on Learning Representations (ICLR 2017), Toulon, France, 2017

    2017

  20. [20]

    Directed graph attention networks for predicting asymmetric drug–drug interactions

    Yi-Yang Feng, Hui Yu, Yue-Hua Feng, Jian-Yu Shi. Directed graph attention networks for predicting asymmetric drug–drug interactions . Briefings in Bioinformatics , V olume 23, Issue 3, May 2022, bbac151. Published: 25 April 2022. arXiv preprint arXiv:2106.07859, 2021

    Pith/arXiv arXiv 2022

  21. [21]

    Qincheng Lu, Jiaqi Zhu, Sitao Luan, Xiao-Wen Chang, Representa- tion Learning on Heterophilic Graph with Directional Neighborhood Attention, arXiv preprint arXiv:2403.01475, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Spatio-temporal Graph Con- volutional Networks: A Deep Learning Framework for Traffic Fore- casting

    Bing Yu, Haoteng Yin, Zhanxing Zhu. Spatio-temporal Graph Con- volutional Networks: A Deep Learning Framework for Traffic Fore- casting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI) , pages 3634–3640, 2018

    2018

  23. [23]

    Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, Huaiyu Wan. Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting. Proceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM), 2020

    2020

  24. [24]

    PeMS: California Freeway Traffic Data

    California Department of Transportation. PeMS: California Freeway Traffic Data. Accessed: 2023-04-28

    2023

  25. [25]

    O. Feldman. The GEH Measure and Quality of the Highway Assign- ment Models. 2012

    2012