INSPIRE-GNN: Intelligent Sensor Placement to Improve Sparse Bicycling Network Prediction via Reinforcement Learning Boosted Graph Neural Networks

Ben Beck; Debjit Bhowmick; Meead Saberi; Mohit Gupta; Rhys Newbury; Shirui Pan

REVIEW 3 major objections 4 minor 63 references

INSPIRE-GNN: Intelligent Sensor Placement to Improve Sparse Bicycling Network Prediction via Reinforcement Learning Boosted Graph Neural Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A reinforcement-learning-boosted graph neural network selects where to place new bicycle counters so that city-wide link-level volume estimates become more accurate than with heuristic or random placement.

desk verdict Plausible applied sensor-placement paper whose central evaluation claim cannot be checked from the abstract alone; the load-bearing question is whether the reported errors are computed on genuinely held-out links. read the letter →

arxiv 2508.00141 v1 pith:6PBI5FWL submitted 2025-07-31 cs.LG cs.AI

Mohit Gupta , Debjit Bhowmick , Rhys Newbury , Meead Saberi , Shirui Pan , Ben Beck This is my paper · ORCID

classification cs.LGcs.AI

keywords sensorplacementbicyclevolumeestimationgraphneuralnetworksreinforcementlearningdeepQ-networksparselink-levelpredictioncyclingnetwork

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

This paper tries to establish that the question of where to place new bicycle count sensors is itself a learnable decision, not a job for generic heuristics. It proposes INSPIRE-GNN, a hybrid graph neural network (GCN plus GAT) whose volume estimator is paired with a Deep Q-Network agent that sequentially picks sensor locations to minimize estimation error. Tested on Melbourne's 15,933-link cycling network with only 141 counters (99 percent sparsity), the claimed result is that RL-chosen locations consistently beat betweenness, closeness, observed activity, and random placement for budgets of 50, 100, 200, and 500 sensors on MSE, RMSE, and MAE. If true, the method gives planners a data-driven way to expand sparse monitoring networks that is more accurate than current heuristic practice.

What carries the argument

The load-bearing object is the pairing of a hybrid Graph Convolutional Network and Graph Attention Network (GCN-GAT) that estimates daily bicycle volumes on every link from the few observed counters, with a Deep Q-Network (DQN) agent that chooses sensor locations sequentially. The GCN-GAT treats the city's 15,933 road segments as graph nodes and propagates the 141 measured volumes across the network; the DQN treats deployment as a Markov decision process in which the state is the current set of observed links, an action adds a new sensor, and the reward reflects the reduction in link-level estimation error. The combined system is what the paper calls INSPIRE-GNN, and the claimed advantage stems from the estimator's error signal steering where new sensors go, rather than from any static centrality measure or observed activity count.

What would settle it

Use a synthetic or fully measured network where true volumes on every link are known, run the DQN agent and the baselines under identical training budgets, then compare prediction error on held-out links; if random or heuristic placement matches the RL-chosen placement over multiple seeds, the claimed advantage is not real.

Watch

Extended reading notes

Core claim

The central discovery claimed is a reinforcement-learning-boosted sensor placement strategy that improves link-level bicycle volume estimation in heavily under-observed networks. On a city-wide road graph where only about one percent of links have counters, the framework treats each additional sensor as an action chosen by a DQN agent, with the reward derived from how much the GNN's volume prediction error drops after the new observation is added. The paper reports that over deployment sizes of 50, 100, 200, and 500 sensors, the RL-selected layouts outperform four common placement baselines (betweenness centrality, closeness centrality, observed bicycling activity, and random placement) across MSE, RMSE, and MAE, and that the framework also beats a set of standard machine learning and deep learning models for the underlying volume-estimation task.

Load-bearing premise

The 141 existing counters are representative enough of all 15,933 road segments that the GNN learns generalizable volume patterns; if the monitored roads are systematically different from unmonitored ones, the estimated volumes and the recommended sensor placements are both skewed.

Editorial extensions

If this is right

At each tested deployment budget, placing sensors where the RL agent recommends reduces link-level volume-estimation error relative to centrality-based, activity-based, and random placement.
The same GNN estimator retrained on RL-chosen sensors outperforms standard machine learning and deep learning baselines for bicycle volume estimation.
Planners can use INSPIRE-GNN's selected locations to expand a sparse sensor network while maximizing the accuracy of city-wide volume estimates.
The framework is reported to work across deployment scales from 50 to 500 new sensors, giving planners a range of investment options.

Reading between the lines

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

Because the reward is estimation-error reduction, the approach is essentially active learning on a graph; a natural test is whether it also beats random placement for other scarce sensor types (pedestrian, traffic, air quality) on road networks.
The claim is evaluated on one city with one topology; a cross-city transfer test would show whether RL placement generalizes or overfits Melbourne's spatial structure.
The 141 existing sensors anchor the entire estimate; leaving out subsets of them and seeing whether recommendations remain stable would quantify how much the advice depends on the initial monitoring layout.

Share X Bluesky LinkedIn Reddit HN

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

Quick take: this is an applied sensor-placement paper for bike volume estimation in Melbourne, and the abstract is the only readable part of the artifact I got. The idea is sensible: train a GNN to estimate link-level volumes, then use a DQN to choose where additional counters should go, and compare against centrality/activity/random baselines. That is a legitimate application of existing methods; not a new principle, but a useful combination, and the 99% sparsity / 141 sensors setup is a real problem worth tackling.

What it does well: the problem is concrete, the baselines are the right ones to beat, and the framing around practical deployment sizes (50–500 sensors) is useful for transport planners. The claim of benchmarking against standard ML/deep models is also appropriate.

Where I have trouble: I can't verify anything beyond the abstract. The full-text extraction I received is mostly replacement characters. More importantly, the abstract reports improvements in MSE/RMSE/MAE but gives no numbers, no error bars, no seeds, and no evaluation protocol. With only 141 monitored segments, the natural test set is tiny and likely biased toward existing counter locations. The DQN reward is the GNN's estimation error, which is fine as an optimization objective, but it means the reported gains over heuristics are only meaningful if the final MSE/RMSE/MAE is evaluated on held-out links that were not used to train the GNN or compute the reward. The abstract doesn't say that. If they evaluated on the same 141 sensors they trained on, the comparison against heuristics is largely circular. This is load-bearing and needs to be checked in the full paper. Also, no mention of how many random seeds or whether the baseline methods were tuned; with small ground truth, the variance could swamp the reported differences.

That said, the stress-test concern about circularity is a request for information, not proof of a flaw. I can't judge the central claim without the full text. If the evaluation is genuinely out-of-sample on unselected segments (or if they use a proper validation scheme on held-out counter locations), this could be a useful paper for the transportation informatics community.

Recommendation: send it to review, but the referee must be asked to verify the evaluation protocol before anything else. If the full paper contains a fair held-out comparison, it's citable for the application; if not, it's another RL/GNN wrapper with no evidence.

Referee Report

3 major / 4 minor

Summary. The paper proposes INSPIRE-GNN, a hybrid framework combining Graph Convolutional Networks, Graph Attention Networks, and a Deep Q-Network-based reinforcement learning agent for optimizing the placement of bicycle count sensors and improving link-level bicycle volume estimation in sparse data environments. The method is applied to Melbourne's bicycling network, which has 15,933 road segments but only 141 count sensors (about 99% sparsity). The abstract claims that the proposed sensor placement strategy outperforms betweenness centrality, closeness centrality, observed bicycling activity, and random placement across deployments of 50, 100, 200, and 500 sensors, with improvements in MSE, RMSE, and MAE, and also benchmarks favorably against standard machine learning and deep learning models. The full text of the submitted artifact is largely corrupted and unreadable, so the review is based primarily on the abstract and the visible fragments of tables and text.

Significance. If the claims are substantiated, the work addresses a practically important problem in sustainable urban transportation: reducing the cost and improving the accuracy of bicycle volume estimation through strategic sensor placement. The central idea of coupling a GNN-based estimation model with an RL agent that selects sensor locations to minimize estimation error is a natural and plausible active-learning formulation. The application to a real city network with very high sparsity is a relevant stress test for the methodology. However, the current submission does not provide the numerical evidence, evaluation protocol, or representativeness analysis needed to assess whether the claims hold, so the significance cannot yet be established.

major comments (3)

[Abstract] The abstract claims 'significant improvements' in MSE, RMSE, and MAE over heuristic placement baselines, but reports no numeric results, no confidence intervals, and no description of the evaluation protocol. This is load-bearing: with only 141 monitored segments (0.9% of the network), the only ground-truth volumes are those 141 sensors, so the comparison must be performed on a held-out subset not used for GNN training or DQN reward computation. The current submission does not demonstrate that this was done, leaving the central claim unverified.
[Abstract (data description)] The paper relies on the representativeness of the 141 existing count sensors for the entire 15,933-segment network. If the monitored segments are concentrated on certain road types or areas, the estimated volumes and optimal placement recommendations will be biased. A coverage analysis (e.g., distribution of road classes, spatial distribution, and comparison with network-wide characteristics) is needed to assess this risk; without it, the claim of whole-network estimation improvement is not justified.
[Evaluation protocol (Tables 4–6)] The visible fragments of Tables 4–6 cannot be read due to corruption in the submitted artifact, so the reported MSE/RMSE/MAE values, the number of random seeds, standard deviations, and the exact train/test/data split for each deployment size cannot be verified. To support the headline claim, the authors must report these details explicitly, including whether the evaluation is out-of-sample with respect to both GNN training and DQN reward computation. Without this, the claimed advantage over baselines could be an artifact of in-sample evaluation.

minor comments (4)

[Abstract] The abstract should report the main quantitative results (e.g., relative reduction in RMSE versus the best baseline) instead of only qualitative claims, to allow readers to immediately assess the magnitude of the claimed improvements.
[Abstract] The manuscript uses both MSE and RMSE as key metrics; since RMSE is the square root of MSE, reporting both is redundant. Consider reporting RMSE and MAE, or explain why both are needed for the comparison.
[Full text] The submitted full text is heavily corrupted with replacement characters, making it impossible to read the methodology, equations, and results. The authors should provide a clean, readable version of the manuscript for review.
[Tables 4–6] The table captions and column headings in Tables 4–6 are not legible; please ensure every table has a clear caption, defines all abbreviations, and states the units and baseline names.

Circularity Check

0 steps flagged · score 0.0 of 10

No exhibited circularity: the RL reward being estimation error is the task objective, and no specific reduction of prediction to fitted input is quotable.

full rationale

The paper's central derivation is an empirical ML pipeline: a GNN volume estimator plus a DQN sensor-placement agent whose reward is defined by estimation error, benchmarked against heuristics on MSE/RMSE/MAE. Optimizing a placement policy against the same metric used to evaluate it is the standard definition of the task, not a circular derivation; circularity would require evidence that the reported metrics are computed on the same sensors used for training or reward while no held-out evaluation exists. The provided full text is heavily corrupted with replacement characters, so the evaluation protocol, data split, and equations cannot be quoted. The abstract alone does not exhibit a reduction of the claimed improvement to its inputs. The limitations section acknowledges that accuracy depends on the representativeness of the 141 existing sensors; this is a generalization-risk concern, not a self-referential derivation. No load-bearing self-citations, imported uniqueness theorems, or ansatz-by-citation steps are identifiable. Accordingly, the honest finding is no significant circularity, with the caveat that the unreadable full text prevents independent verification of the held-out evaluation protocol.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The model relies on standard ML assumptions and the fidelity of the sensor data. No new physical entities are proposed.

free parameters (2)

GNN hyperparameters (number of layers, hidden dimensions, attention units)
Not specified in the abstract; these are likely tuned on the Melbourne dataset and directly affect reported performance.
DQN hyperparameters (learning rate, exploration rate, replay buffer size, reward scaling)
Not specified in the abstract; RL training dynamics depend on these choices.

assumptions (2)

domain assumption The road network graph structure captures bicycle volume correlations between nearby segments.
The GCN/GAT architecture assumes that neighboring road segments have related ridership, which is the basis for propagating information from counted to uncounted segments.
domain assumption The observed bicycle counts from existing sensors are accurate ground truth.
All training and evaluation uses these counts as labels; systematic sensor error would bias the model and the placement decisions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of INSPIRE-GNN: Intelligent Sensor Placement to Improve Sparse Bicycling Network Prediction via Reinforcement Learning Boosted Graph Neural Networks." pith.science (2026). https://pith.science/paper/6PBI5FWL

@misc{pith2026250800141,
  author       = {Pith},
  title        = {Pith review of: INSPIRE-GNN: Intelligent Sensor Placement to Improve Sparse Bicycling Network Prediction via Reinforcement Learning Boosted Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6PBI5FWL}},
  note         = {Machine review of arXiv:2508.00141}
}

read the original abstract

Accurate link-level bicycling volume estimation is essential for sustainable urban transportation planning. However, many cities face significant challenges of high data sparsity due to limited bicycling count sensor coverage. To address this issue, we propose INSPIRE-GNN, a novel Reinforcement Learning (RL)-boosted hybrid Graph Neural Network (GNN) framework designed to optimize sensor placement and improve link-level bicycling volume estimation in data-sparse environments. INSPIRE-GNN integrates Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) with a Deep Q-Network (DQN)-based RL agent, enabling a data-driven strategic selection of sensor locations to maximize estimation performance. Applied to Melbourne's bicycling network, comprising 15,933 road segments with sensor coverage on only 141 road segments (99% sparsity) - INSPIRE-GNN demonstrates significant improvements in volume estimation by strategically selecting additional sensor locations in deployments of 50, 100, 200 and 500 sensors. Our framework outperforms traditional heuristic methods for sensor placement such as betweenness centrality, closeness centrality, observed bicycling activity and random placement, across key metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). Furthermore, our experiments benchmark INSPIRE-GNN against standard machine learning and deep learning models in the bicycle volume estimation performance, underscoring its effectiveness. Our proposed framework provides transport planners actionable insights to effectively expand sensor networks, optimize sensor placement and maximize volume estimation accuracy and reliability of bicycling data for informed transportation planning decisions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 55 canonical work pages

[1]

Atap guidelines: National guidelines for transport system management in australia

Australian Transport Assessment and Planning (2022). Atap guidelines: National guidelines for transport system management in australia. Available at: https://www.atap.gov.au/

work page 2022
[2]

Ba, J. L. (2016). Layer normalization. arXiv preprint arXiv:1607.06450

arXiv 2016
[3]

Z., Seneviratne, S., Pettit, C., Vu, H

Bhowmick, D., Saberi, M., Stevenson, M., Thompson, J., Winters, M., Nelson, T., Leao, S. Z., Seneviratne, S., Pettit, C., Vu, H. L., et al. (2023). A systematic scoping review of methods for estimating link-level bicycling volumes. Transport reviews , 43(4):622--651

work page 2023
[4]

and Dill, J

Buehler, R. and Dill, J. (2016). Bikeway networks: A review of effects on cycling. Transport reviews , 36(1):9--27

work page 2016
[5]

and Pucher, J

Buehler, R. and Pucher, J. (2012). Cycling to work in 90 large american cities: new evidence on the role of bike paths and lanes. Transportation , 39:409--432

work page 2012
[6]

Chen, C., Wei, H., Xu, N., Zheng, G., Yang, M., Xiong, Y., Xu, K., and Li, Z. (2020). Toward a thousand lights: Decentralized deep reinforcement learning for large-scale traffic signal control. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 3414--3421

work page 2020
[7]

Cooper, C. H. (2017). Using spatial network analysis to model pedal cycle flows, risk and mode choice. Journal of transport geography , 58:157--165

work page 2017
[8]

Darvariu, V.-A., Hailes, S., and Musolesi, M. (2024). Graph reinforcement learning for combinatorial optimization: A survey and unifying perspective. arXiv preprint arXiv:2404.06492

arXiv 2024

Show all 63 references

[9]

De Geus, B., De Smet, S., Nijs, J., and Meeusen, R. (2007). Determining the intensity and energy expenditure during commuter cycling. British journal of sports medicine , 41(1):8--12

2007
[10]

J., Boogaard, H., Nijland, H., and Hoek, G

De Hartog, J. J., Boogaard, H., Nijland, H., and Hoek, G. (2010). Do the health benefits of cycling outweigh the risks? Environmental health perspectives , 118(8):1109--1116

2010
[11]

Elassy, M., Al-Hattab, M., Takruri, M., and Badawi, S. (2024). Intelligent transportation systems for sustainable smart cities. Transportation Engineering , page 100252

2024
[12]

Cycling remains at high levels in 2023 edition of EuroVelo Usage Monitoring Report

EuroVelo (2024). Cycling remains at high levels in 2023 edition of EuroVelo Usage Monitoring Report . https://pro.eurovelo.com/news/\\(Report ID: 2024-03-11). Accessed: 2025-04-24

2024
[13]

Fakhouri, A. A. and Soltani, R. (2020). Multi-objective robust optimization for the traffic sensors location problem. IEEE Access , 9:6225--6238

2020
[14]

Ferster, C., Nelson, T., Laberee, K., and Winters, M. (2021). Mapping bicycling exposure and safety risk using strava metro. Applied Geography , 127:102388

2021
[15]

Fischer, J., Nelson, T., and Winters, M. (2020). Comparing spatial associations of commuting versus recreational ridership captured by the strava fitness app. Findings

2020
[16]

Freeman, L. (1977). A set of measures of centrality based on betweenness. Sociometry

1977
[17]

Fu, H., Ma, H., Liu, Y., and Lu, D. (2016). A vehicle classification system based on hierarchical multi-svms in crowded traffic scenes. Neurocomputing , 211:182--190

2016
[18]

Gao, J., Shen, Y., Liu, J., Ito, M., and Shiratori, N. (2017). Adaptive traffic signal control: Deep reinforcement learning algorithm with experience replay and target network. arXiv preprint arXiv:1705.02755

2017 arXiv
[19]

L., Spak, S

Grabow, M. L., Spak, S. N., Holloway, T., Stone Jr, B., Mednick, A. C., and Patz, J. A. (2012). Air quality and exercise-related health benefits from reduced car travel in the midwestern united states. Environmental health perspectives , 120(1):68--76

2012
[20]

B., Medury, A., and Schneider, R

Griswold, J. B., Medury, A., and Schneider, R. J. (2011). Pilot models for estimating bicycle intersection volumes. Transportation research record , 2247(1):1--7

2011
[21]

Gu, Y., Qian, Z., and Zhang, G. (2017). Traffic state estimation for urban road networks using a link queue model. Transportation research record , 2623(1):29--39

2017
[22]

Guide, T. M. (2001). Traffic monitoring guide. Federal Highway Administration

2001
[23]

Gupta, M., Bhowmick, D., Saberi, M., Pan, S., and Beck, B. (2024). Evaluating the effects of data sparsity on the link-level bicycling volume estimation: A graph convolutional neural network approach. arXiv preprint arXiv:2410.08522

2024
[24]

and Y lmaz, Y

Haydari, A. and Y lmaz, Y. (2020). Deep reinforcement learning for intelligent transportation systems: A survey. IEEE Transactions on Intelligent Transportation Systems , 23(1):11--32

2020
[25]

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 770--778

2016
[26]

Jestico, B., Nelson, T., and Winters, M. (2016). Mapping ridership using crowdsourced cycling data. Journal of transport geography , 52:90--97

2016
[27]

A., Mills, M

Klein, L. A., Mills, M. K., Gibson, D. R., et al. (2006). Traffic detector handbook: Volume i. Technical report, Turner-Fairbank Highway Research Center

2006
[28]

S., and Dhakad, N

Kumar, N., Rahman, S. S., and Dhakad, N. (2020). Fuzzy inference enabled deep reinforcement learning-based traffic light control for intelligent transportation system. IEEE Transactions on Intelligent Transportation Systems , 22(8):4919--4928

2020
[29]

Kwigizile, V., Oh, J.-S., Kwayu, K., et al. (2019). Integrating crowdsourced data with traditionally collected data to enhance estimation of bicycle exposure measure. Technical report, Western Michigan University. Transportation Research Center for Livable …

2019
[30]

and Sener, I

Lee, K. and Sener, I. N. (2021). Strava metro data for bicycle monitoring: a literature review. Transport reviews , 41(1):27--47

2021
[31]

Li, Y., Yu, R., Shahabi, C., and Liu, Y. (2017). Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926

2017 arXiv
[32]

and Fan, W

Lin, Z. and Fan, W. D. (2020). Modeling bicycle volume using crowdsourced data from strava smartphone application. International journal of transportation science and technology , 9(4):334--343

2020
[33]

Lindsay, G., Macmillan, A., and Woodward, A. (2011). Moving urban trips from cars to bicycles: impact on health and emissions. Australian and New Zealand journal of public health , 35(1):54--60

2011
[34]

Livingston, M., McArthur, D., Hong, J., and English, K. (2021). Predicting cycling volumes using crowdsourced activity data. Environment and Planning B: Urban Analytics and City Science , 48(5):1228--1244

2021
[35]

May, M., Hecker, D., K \"o rner, C., Scheider, S., and Schulz, D. (2008). A vector-geometry based spatial knn-algorithm for traffic frequency predictions. In 2008 IEEE International Conference on Data Mining Workshops , pages 442--447. IEEE

2008
[36]

Meyer, M. D. (2016). Transportation planning handbook . John Wiley & Sons

2016
[37]

M., Hyun, K

Miah, M. M., Hyun, K. K., Mattingly, S. P., Broach, J., McNeil, N., and Kothuri, S. (2022). Challenges and opportunities of emerging data sources to estimate network-wide bike counts. Journal of transportation engineering, Part A: Systems , 148(3):04021122

2022
[38]

M., Hyun, K

Miah, M. M., Hyun, K. K., Mattingly, S. P., and Khan, H. (2023). Estimation of daily bicycle traffic using machine and deep learning techniques. Transportation , 50(5):1631--1684

2023
[39]

A., Ferster, C., Roy, A., and Winters, M

Nelson, T. A., Ferster, C., Roy, A., and Winters, M. (2023). Bicycle streetscapes: a data driven approach to mapping streets based on bicycle usage. International Journal of Sustainable Transportation , 17(8):931--941

2023
[40]

Newman, M. E. (2005). A measure of betweenness centrality based on random walks. Social networks , 27(1):39--54

2005
[41]

OpenStreetMap

OpenStreetMap contributors (2022). OpenStreetMap . https://www.openstreetmap.org

2022
[42]

Gajewski, Hołyst, J

Paluch, R., Łukasz G. Gajewski, Hołyst, J. A., and Szymanski, B. K. (2020). Optimizing sensors placement in complex networks for localization of hidden signal source: A review. Future Generation Computer Systems , 112:1070--1092

2020
[43]

and Liu, S

Pan, W. and Liu, S. Q. (2023). Deep reinforcement learning for the dynamic and uncertain vehicle routing problem. Applied Intelligence , 53(1):405--422

2023
[44]

Pearson, L., Bhowmick, D., Winters, M., Saberi, M., Nelson, T., Pettit, C., Nice, K., Dai, D., Gupta, M., and Beck, B. (2025). Gender differences in bicycle infrastructure use and preferences: A disconnect between ideals and reality. Available at SSRN 5108271

2025
[45]

and Buehler, R

Pucher, J. and Buehler, R. (2008). Cycling for everyone: lessons from europe. Transportation research record , 2074(1):58--65

2008
[46]

Rahmani, S., Baghbani, A., Bouguila, N., and Patterson, Z. (2023). Graph neural networks for intelligent transportation systems: A survey. IEEE Transactions on Intelligent Transportation Systems , 24(8):8846--8885

2023
[47]

C., Harris, M

Reynolds, C. C., Harris, M. A., Teschke, K., Cripton, P. A., and Winters, M. (2009). The impact of transportation infrastructure on bicycling injuries and crashes: a review of the literature. Environmental health , 8:1--19

2009
[48]

Salon, D. (2016). Estimating pedestrian and cyclist activity at the neighborhood scale. Journal of transport geography , 55:11--21

2016
[49]

Schoner, J., Proulx, F., Knapp de Orva \ n anos, K., and Almdale, B. (2021). Prioritizing pedestrian and bicyclist count locations for volume estimation. Transportation research record , 2675(10):277--290

2021
[50]

Senturk, I. F. and Akkaya, K. (2014). Connectivity restoration in disjoint wireless sensor networks using centrality measures. In 39th Annual IEEE Conference on Local Computer Networks Workshops , pages 616--622

2014
[51]

Strava Metro

Strava Metro contributors (2022). Strava Metro . https://metro.strava.com

2022
[52]

Tokic, M. (2010). Adaptive -greedy exploration in reinforcement learning based on value differences. In Annual conference on artificial intelligence , pages 203--210. Springer

2010
[53]

Velickovic, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y., et al. (2017). Graph attention networks. stat , 1050(20):10--48550

2017
[54]

Wang, T., Zhu, Z., Zhang, J., Tian, J., and Zhang, W. (2024). A large-scale traffic signal control algorithm based on multi-layer graph deep reinforcement learning. Transportation Research Part C: Emerging Technologies , 162:104582

2024
[55]

Wang, Z., Li, H.-X., and Chen, C. (2019). Reinforcement learning-based optimal sensor placement for spatiotemporal modeling. IEEE transactions on cybernetics , 50(6):2861--2871

2019
[56]

Wen, L. M. and Rissel, C. (2008). Inverse associations between cycling to work, public transport, and overweight and obesity: findings from a population based study in australia. Preventive medicine , 46(1):29--32

2008
[57]

and Branion-Calles, M

Winters, M. and Branion-Calles, M. (2017). Cycling safety: Quantifying the under reporting of cycling incidents in vancouver, british columbia. Journal of Transport & Health , 7:48--53

2017
[58]

Winters, M., Teschke, K., Brauer, M., and Fuller, D. (2016). Bike score : Associations between urban bikeability and cycling behavior in 24 cities. International journal of behavioral nutrition and physical activity , 13:1--10

2016
[59]

Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip, S. Y. (2020). A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems , 32(1):4--24

2020
[60]

Yu, B., Yin, H., and Zhu, Z. (2017). Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. Arxiv Preprint Arxiv:1709.04875

2017 arXiv
[61]

Zhang, J., Shi, X., Xie, J., Ma, H., King, I., and Yeung, D.-Y. (2018). Gaan: Gated attention networks for learning on large and spatiotemporal graphs. arXiv preprint arXiv:1803.07294

2018 arXiv
[62]

Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., Deng, M., and Li, H. (2019). T-gcn: A temporal graph convolutional network for traffic prediction. IEEE transactions on intelligent transportation systems , 21(9):3848--3858

2019
[63]

Zhao, S., Zhao, P., and Cui, Y. (2017). A network centrality measure framework for analyzing urban traffic flow: A case study of wuhan, china. Physica A: Statistical Mechanics and its Applications , 478:143--157

2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Atap guidelines: National guidelines for transport system management in australia

Australian Transport Assessment and Planning (2022). Atap guidelines: National guidelines for transport system management in australia. Available at: https://www.atap.gov.au/

work page 2022

[2] [2]

Ba, J. L. (2016). Layer normalization. arXiv preprint arXiv:1607.06450

arXiv 2016

[3] [3]

Z., Seneviratne, S., Pettit, C., Vu, H

Bhowmick, D., Saberi, M., Stevenson, M., Thompson, J., Winters, M., Nelson, T., Leao, S. Z., Seneviratne, S., Pettit, C., Vu, H. L., et al. (2023). A systematic scoping review of methods for estimating link-level bicycling volumes. Transport reviews , 43(4):622--651

work page 2023

[4] [4]

and Dill, J

Buehler, R. and Dill, J. (2016). Bikeway networks: A review of effects on cycling. Transport reviews , 36(1):9--27

work page 2016

[5] [5]

and Pucher, J

Buehler, R. and Pucher, J. (2012). Cycling to work in 90 large american cities: new evidence on the role of bike paths and lanes. Transportation , 39:409--432

work page 2012

[6] [6]

Chen, C., Wei, H., Xu, N., Zheng, G., Yang, M., Xiong, Y., Xu, K., and Li, Z. (2020). Toward a thousand lights: Decentralized deep reinforcement learning for large-scale traffic signal control. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 3414--3421

work page 2020

[7] [7]

Cooper, C. H. (2017). Using spatial network analysis to model pedal cycle flows, risk and mode choice. Journal of transport geography , 58:157--165

work page 2017

[8] [8]

Darvariu, V.-A., Hailes, S., and Musolesi, M. (2024). Graph reinforcement learning for combinatorial optimization: A survey and unifying perspective. arXiv preprint arXiv:2404.06492

arXiv 2024