Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Commute Networks as a Signature of Urban Socioeconomic Performance: Evaluating Mobility Structures with Deep Learning Models

T0 review · 4 major / 5 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 Commute network structure alone predicts median household income at the census-tract level in most U.S. cities tested; a graph neural network pipeline does it better than feature-based models.

desk verdict A credible extension of the authors' earlier commute-network work, but the abstract's core claim—network structure alone, without node features—is not actually established because spatial coordinates are used as an input and no spatial-only or null-graph baseline is provided. read the letter →

arxiv 2507.04027 v1 pith:6TSQAS7J submitted 2025-07-05 cs.LG

Devashish Khulbe , Alexander Belyi , Stanislav Sobolevsky This is my paper
classification cs.LG
keywords commutenetworksmobilitygraphneuralsocioeconomicmodelingmedianincomepredictionurbannetworkembeddingcensustracts
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

The paper sets out to show that who commutes where inside a city is itself a socioeconomic signal. Using census origin-destination commute records to build a network whose nodes are census tracts, and giving the model no other node information, the authors report that this structure alone predicts median household income out-of-sample with R2 around 0.5 to 0.7 in most of 12 U.S. cities, reaching a best of 0.75 in San Jose. To do this they introduce a single end-to-end pipeline that stacks graph-convolution or graph-attention layers with a feedforward network, replacing the usual two-stage approach of first learning a network embedding and then regressing on it. If the claim holds, urban researchers can model neighborhood income from mobility patterns alone, without waiting for periodic demographic surveys, and planners can ask how a proposed transit change would alter predicted income by editing the network.

What carries the argument

The central object is the commute network graph, whose nodes are census tracts and whose edge weights are the daily counts of people commuting between tracts. The load-bearing mechanism is message passing: each graph-convolution or graph-attention layer replaces a tract's representation with an aggregation of its neighbors' representations, so the final two-layer embedding encodes each tract's position in the city's mobility structure, and a final feedforward network maps that embedding to median income. Because all parameters are trained jointly on the income objective, the learned embedding is shaped specifically for socioeconomic prediction rather than for generic network reconstruction. The separate baseline embedding model works by training a three-layer feedforward network to reconstruct the full adjacency matrix, with the node vectors themselves treated as trainable parameters.

What would settle it

Run the same GNN+VNN training on the same cities but, for each test tract, delete all edges connecting it to training tracts before message passing; if the out-of-sample R2 falls to the level of local-feature-only baselines, the reported network signal is mostly spatial leakage, whereas if it stays near 0.5 to 0.75, the structural claim is supported.

Watch

Extended reading notes

Core claim

The central discovery claimed is that the structure of a commute network, without any census-tract attributes as input, is a strong predictor of a tract's median household income. Across 12 U.S. cities, a two-layer graph-convolution or graph-attention network followed by a feedforward mapping achieves out-of-sample R2 values mostly in the 0.5 to 0.7 range, with the best single result of 0.75 in San Jose using an SVD-based network embedding. The proposed end-to-end GNN+VNN pipeline generally outperforms models trained on 311-complaint signatures and population or job density, and it also tends to beat the separate embedding-plus-regression approach, supporting the paper's claim that network topology alone carries socioeconomic information.

Load-bearing premise

The experiment assumes that training on the full commute graph does not leak income information into the held-out tracts, even though neighboring tracts tend to have similar incomes and the model's message passing reaches across the train/test split.

Editorial extensions

If this is right

  • If the claim is correct, census-tract income can be estimated from commute structure alone, which matters when income data are sparse, delayed, or unavailable.
  • The end-to-end GNN+VNN pipeline removes the need for a separately trained embedding model, simplifying deployment and aligning representation learning with the prediction target.
  • Network-only models outperform 311-complaint and density-based feature models in most of the 12 cities, suggesting that inter-neighborhood topology captures information that local features miss.
  • City-level variation, with weak results in Los Angeles and Phoenix, indicates that the signal depends on urban form, so mobility-based inference will not transfer uniformly across metropolitan areas.
  • Adding 311-complaint features on top of network embeddings improved results in New York and Chicago, so network structure and local context appear complementary rather than redundant.

Reading between the lines

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

  • Extension: The paper does not test for spatial leakage across the train/test split; masking all messages that pass between training and test tracts would show whether the reported R2 is genuine network structure or simply spatial smoothing of strongly autocorrelated incomes.
  • Extension: Because income is known to shape commute distances and modal choices, comparing the model against a simple gravity-model baseline would isolate how much of the predictive power comes from the network topology rather than from known income-mobility correlations.
  • Extension: The same pipeline could be retrained on faster-updating mobility data, such as mobile-phone or transit records, to produce near-real-time socioeconomic indicators; the paper mentions this possibility but does not test it.
  • Extension: The framework is transferable in principle to other socioeconomic outcomes and to non-U.S. cities, but only where commute records have comparable tract-level granularity and where the urban form resembles the cities studied.
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.

Referee Report

4 major / 5 minor

Summary. The paper constructs commute-flow networks between census tracts from LEHD data for 12 U.S. cities and uses them to predict tract-level median household income. Three modeling strategies are compared: (i) VNN-based embeddings learned by reconstructing the adjacency matrix and then used as inputs to a supervised MLP; (ii) an end-to-end GCN+VNN pipeline; and (iii) a GAT+VNN variant. The inputs to these models are initial node representations taken from spatial coordinates, SVD, Laplacian eigenvectors, or random-walk (PageRank) embeddings. The authors report out-of-sample R² values under a 70:30 node split, compare against an MLP trained on 311 complaint features, and conclude that mobility network structure alone, without node features, provides substantial predictive power and outperforms conventional machine learning baselines.

Significance. If the central claim were established, the paper would make a useful empirical contribution: commute-flow networks are publicly available, regularly updated, and could serve as a scalable signal for urban socioeconomic modeling, complementing or replacing expensive local feature engineering. The use of 12 cities, a public data repository, and an examination of several embedding schemes are strengths. However, the headline claims are currently not supported by the experimental design. The best-performing input is explicitly spatial coordinates, which are node features; all graph models use the full adjacency structure during training while the supervised split is on nodes; and no spatial-only or null-graph baseline is provided. These confounds mean the paper does not yet isolate 'network structure' as the source of predictive performance. The contribution is therefore promising but requires substantial additional experimentation and reframing.

major comments (4)
  1. [§2.2.2, §2.3, Eqs. (1)-(6)] The out-of-sample evaluation is transductive and may inflate R² through spatial smoothing. The VNN embedding is trained on the full adjacency matrix before the supervised split, and the GCN/GAT layers in Eqs. (3)-(6) aggregate over the complete neighborhood N(i), which includes both training and test tracts, while the supervised split is 70:30 on nodes (§2.3). Because median income is strongly spatially autocorrelated, message passing can propagate income-correlated information from training tracts to test tracts even if test labels are never used. The paper provides no control such as an edge-level split, an inductive evaluation, or a version of the model in which test nodes are prevented from attending to training nodes. This concern is load-bearing for the claim that the reported R² values reflect genuine out-of-sample predictive skill.
  2. [§2.2.1, §3, Table 2] The abstract's claim that prediction is achieved 'without considering any node features' is contradicted by the paper's own best-performing configuration. Section 2.2.1 lists 'spatial embedding (location of regions)' as an initial input, and Table 2 shows that this coordinate-based input is the best or near-best configuration in NYC, Chicago, Boston, and elsewhere. Spatial coordinates are node-level features. To support the isolation claim, the authors need a spatial-only baseline (e.g., an MLP or kriging on coordinates, or a graph Laplacian smoother applied to training incomes) and a direct comparison showing the incremental value of the network topology beyond the coordinate signal. Without this, the central 'network structure alone' claim is not established.
  3. [§3, Table 2] There is no null-graph or topology-control experiment. The reported results do not separate the effect of network topology from the effect of geographic proximity or from edge weights that are themselves spatial. A degree-preserving random rewiring of the commute graph, or a model using only node degrees as features, would provide a minimal control. Similarly, comparing the GNN+VNN model to a standard label-propagation or spatial-smoothing baseline would clarify whether the improvement over an MLP on 311 features comes from relational structure or simply from nearby tracts having similar incomes. Without such controls, the conclusion that 'mobility network structures provide significant predictive performance' is not supported by the experimental evidence.
  4. [§3, Table 2, Appendix A.3, Table S1] The reported evaluation overstates the reliability of the headline numbers. Table 2 gives point estimates for each city without error bars for the GNN/GAT results, and the best configuration is selected per city across four embedding types and, in Appendix A.3, across multiple embedding dimensions. Appendix Table S1 shows substantial variation with embedding dimension (e.g., NYC R² ranges from 0.40 to 0.53; San Jose from 0.47 to 0.75). The paper should report mean and standard deviation over multiple random seeds, present all embedding dimensions rather than a per-city optimum, and justify the choice of a single 311-feature MLP as the 'conventional machine learning' baseline before claiming superiority over previous models.
minor comments (5)
  1. [§4.1] The Discussion states that spatial embeddings 'may not directly represent network structure' while the Abstract claims the models work 'without considering any node features'; this tension should be resolved explicitly in the main text.
  2. [§2.2.3] The text says the MSE objective from Eq. (2) is changed to '1/N^2 sum_{i=1}^N (Y_i - Y_hat_i)^2', but the sum is over N, not N^2; the normalization and indexing should be corrected.
  3. [Table 2, Appendix A.3] Table 2 does not indicate which embedding dimension was used for each city, making it impossible to connect the headline results to the stability analysis in Table S1; a column or appendix note should specify the dimension per city.
  4. [Appendix A.1] The appendix states that 'mobility information for Chicago and Boston was retrieved from the LEHD' but Table 1 covers 12 cities; the data source description should cover all cities.
  5. [Data Availability] Only the data are deposited; training code, model configurations, and random seeds are not provided, which limits reproducibility of the exact R² values in Table 2.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; the prediction pipeline is empirically self-contained. Self-citations are contextual, and the 'no node features' claim is a construct-validity concern rather than a circular step.

full rationale

The paper does not present a derivation that reduces to its inputs. The VNN embeddings are trained to reconstruct the commute adjacency matrix without using income, and the subsequent income model is fit and evaluated on held-out nodes. The GNN+VNN pipeline is trained end-to-end on income, but the reported R2 values are out-of-sample on a 70:30 node split, so the predictions are not forced by construction. The self-citations ([24], [40], [41]) motivate the problem and cite prior evidence that mobility correlates with income, but they are not used to generate the numerical results, which come from the paper's own experiments against the 311-feature baseline. A caveat worth noting, though it is not circularity: the abstract claims prediction 'without considering any node features,' yet Section 2.2.1 lists 'spatial embedding (location of regions)' among the initial inputs, and Table 2 shows spatial embeddings are often the best configuration. This is an internal inconsistency and a threat to the construct validity of the 'network structure only' claim, especially without a spatial-only baseline or a null-graph control. But this is a correctness/confounding issue, not a circular reduction: the models are genuinely fit and evaluated out-of-sample, and no fitted parameter is renamed as a prediction.

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

The central claims rest on the mobility data being a faithful representation of urban interaction, on the choice of income as the target, and on the experimental protocol (transductive use of the graph, per-city selection of embedding dimension). The 'no node features' narrative depends on a definitional choice to treat spatial coordinates as an embedding rather than a feature. These are assumptions rather than derived laws, and none are machine-checked.

free parameters (5)
  • embedding dimension d = 5 for most cities, 10 for NYC, San Jose, Phoenix (Appendix Table S1)
    Chosen per city based on validation R²; affects all reported results.
  • GNN depth = 2 layers
    Standard choice (Kipf & Welling), not tuned; central to the GNN+VNN architecture.
  • VNN hidden layer sizes = (32,64,32) for supervision, (4d,3d,d) for the embedding model
    Hand-chosen, not justified by data or theory.
  • 70:30 split seed = not reported
    No seed or repeated splits reported, so variance across splits is unknown.
  • GAT attention heads K = not reported
    Unspecified number of heads in Equation 6 multi-head attention.
assumptions (5)
  • domain assumption LEHD LODES commute flows represent the day-time mobility network at census-tract level
    Section 2.1; if flows are noisy or unrepresentative, the network is mis-specified.
  • domain assumption Median household income is an adequate socioeconomic indicator for this task
    Section 2.3; the paper acknowledges income is not complete.
  • domain assumption 311 complaint proportions are a valid baseline for urban socioeconomic modeling
    Section 2.1 and reference [8]; comparison depends on this.
  • ad hoc to paper Spatial coordinates can be treated as an 'embedding' and not as node features for the 'no node features' claim
    Section 2.2.1; spatial embedding is the strongest configuration, so the claim of no node features is misleading.
  • ad hoc to paper The full graph can be used in training (transductive) without leaking the target into test predictions
    Section 2.3; no discussion of spatial autocorrelation leakage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Commute Networks as a Signature of Urban Socioeconomic Performance: Evaluating Mobility Structures with Deep Learning Models." pith.science (2026). https://pith.science/paper/6TSQAS7J

@misc{pith2026250704027,
  author       = {Pith},
  title        = {Pith review of: Commute Networks as a Signature of Urban Socioeconomic Performance: Evaluating Mobility Structures with Deep Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6TSQAS7J}},
  note         = {Machine review of arXiv:2507.04027}
}
read the original abstract

Urban socioeconomic modeling has predominantly concentrated on extensive location and neighborhood-based features, relying on the localized population footprint. However, networks in urban systems are common, and many urban modeling methods don't account for network-based effects. In this study, we propose using commute information records from the census as a reliable and comprehensive source to construct mobility networks across cities. Leveraging deep learning architectures, we employ these commute networks across U.S. metro areas for socioeconomic modeling. We show that mobility network structures provide significant predictive performance without considering any node features. Consequently, we use mobility networks to present a supervised learning framework to model a city's socioeconomic indicator directly, combining Graph Neural Network and Vanilla Neural Network models to learn all parameters in a single learning pipeline. Our experiments in 12 major U.S. cities show the proposed model outperforms previous conventional machine learning models. This work provides urban researchers methods to incorporate network effects in urban modeling and informs stakeholders of wider network-based effects in urban policymaking and planning.

Figures

Figures reproduced from arXiv: 2507.04027 by the authors.

Figure 1
Figure 1. Median income distribution across NYC census tracts [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Clustering of census tracts in cities based on SVD embedding of mobility networks – spatial patterns reveal distinction of high-income from low-income neighborhoods a VNN to reconstruct the edges of the mobility network, thereby learning structural representations of nodes. This training is conducted in a self-supervised manner, where the objective is to reconstruct the original adjacency structure of the network. T… view at source ↗
Figure 3
Figure 3. VNN-based embedding model – Socioeconomic modeling is done in a two-step process: 1. Trainable regions’ embeddings are processed and fed to an MLP to reconstruct the original O-D mobility flow matrix, 2. The learned embeddings are then fed as input to an MLP for modeling median income The trained matrix E can thus be regarded as an embedding of the network nodes. Each d-dimensional vector within this embedding corre… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Graph-based models – GCN/GAT layers are stacked with the MLP layers to directly model the socioeco￾nomic feature We consider a 2 layer GNN model, considering the network adjacency matrix. A two-layer GNN architecture has been widely recognized as effective for various …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Urban delineation through the lens of commute networks: Leveraging graph embeddings to distinguish socioeconomic groups in cities

    cs.SI 2025-07 conditional novelty 5.0 of 10

    A self-supervised GNN trained on census commute flows yields urban community delineations that match modularity-based methods in spatial coherence and reveal income-segregated neighborhoods in 12 U.S. metro areas.

Reference graph

Works this paper leans on

54 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    Rosvall, A

    M. Rosvall, A. Trusina, P. Minnhagen, and K. Sneppen. Networks and cities: An information perspective. Physical Review Letters, 94(2):028701, 2005. 14 A PREPRINT - JULY 17, 2025 Figure A4: Clustering of Laplacian embedding of mobility networks - while results are not so spatially cohesive as SVD and Pagerank methods, we can still distinguish high-income n...

    work page 2005

  2. [2]

    Pflieger and C

    G. Pflieger and C. Rozenblat. Introduction. urban networks and network theory: the city as the connector of multiple networks. Urban Studies, 47(13):2723–2735, 2010

    work page 2010

  3. [3]

    Jiang and C

    B. Jiang and C. Claramunt. Topological analysis of urban street networks.Environment and Planning B: Planning and design, 31(1):151–162, 2004

    work page 2004

  4. [4]

    X. Li, Z. Lv, Z. Zheng, C. Zhong, I. H. Hijazi, and S. Cheng. Assessment of lively street network based on geographic information system and space syntax. Multimedia Tools and Applications, 76:17801–17819, 2017

    work page 2017

  5. [5]

    Urban road network expansion and its driving variables: A case study of nanjing city

    Ge Shi, Jie Shan, Liang Ding, Peng Ye, Yang Li, and Nan Jiang. Urban road network expansion and its driving variables: A case study of nanjing city. International Journal of Environmental Research and Public Health , 16(13):2318, 2019

    work page 2019

  6. [6]

    Y . Xu, A. Belyi, I. Bojic, and C. Ratti. Human mobility and socioeconomic status: analysis of singapore and boston. Computers, Environment and Urban Systems, 72:51–67, 2018

    work page 2018

  7. [7]

    Lee and J

    S. Lee and J. Lin. Natural amenities, neighbourhood dynamics, and persistence in the spatial distribution of income. Review of Economic Studies, 85(1):663–694, 2018

    work page 2018

  8. [8]

    L. Wang, C. Qian, P. Kats, C. Kontokosta, and S. Sobolevsky. Structure of 311 service requests as a signature of urban location. PLoS ONE, 12(10):e0186314, 2017

    work page 2017

Show all 54 references
  1. [9]

    node2vec: Scalable feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, page 855–864, 2016. 15 A PREPRINT - JULY 17, 2025

    2016

  2. [10]

    Line: Large-scale information network embedding

    Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web , page 1067–1077, 2015

    2015

  3. [11]

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

    2016 arXiv

  4. [12]

    Sobolevsky and A

    S. Sobolevsky and A. Belyi. Graph neural network inspired algorithm for unsupervised network community detection. Applied Network Science, 7:63, 2022

    2022

  5. [13]

    Velickovic, G

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

    2017

  6. [14]

    Kempinska and R

    K. Kempinska and R. Murcio. Modelling urban networks using variational autoencoders. Applied Network Science, 4:114, 2019

    2019

  7. [15]

    Pagani, A

    A. Pagani, A. Mehrotra, and M. Musolesi. Graph input representations for machine learning applications in urban network analysis. Environment and Planning B: Urban Analytics and City Science, 48(4):741–758, 2021

    2021

  8. [16]

    Huang, D

    W. Huang, D. Zhang, G. Mai, X. Guo, and L. Cui. Learning urban region representations with pois and hierar- chical graph infomax. ISPRS Journal of Photogrammetry and Remote Sensing, 196:134–145, 2023

    2023

  9. [17]

    Kim and Y

    N. Kim and Y . Yoon. Effective urban region representation learning using heterogeneous urban graph attention network (hugat). arXiv preprint arXiv:2202.09021, 2022

    2022 arXiv

  10. [18]

    Mishina et al

    M. Mishina et al. Prediction of urban population-facilities interactions with graph neural network. In Com- putational Science and Its Applications – ICCSA 2023 , volume 13956 of Lecture Notes in Computer Science . Springer, Cham, 2023

    2023

  11. [19]

    X. Yan, T. Ai, M. Yang, and H. Yin. A graph convolutional neural network for classification of building patterns using spatial vector data. In ISPRS journal of photogrammetry and remote sensing, volume 150, pages 259–273, 2019

    2019

  12. [20]

    Y . Xu, S. Jin, Z. Chen, X. Xie, S. Hu, and Z. Xie. Application of a graph convolutional network with visual and semantic features to classify urban scenes. International Journal of Geographical Information Science , 36(10):2009–2034, 2022

    2009

  13. [21]

    Graph neural network approach to predict the effects of road capacity reduction policies: A case study for paris, france, 2024

    Elena Natterer, Roman Engelhardt, Sebastian Hörl, and Klaus Bogenberger. Graph neural network approach to predict the effects of road capacity reduction policies: A case study for paris, france, 2024

    2024

  14. [22]

    A multi-modal graph neural network approach to traffic risk forecasting in smart urban sensing

    Yang Zhang, Xiangyu Dong, Lanyu Shang, Daniel Zhang, and Dong Wang. A multi-modal graph neural network approach to traffic risk forecasting in smart urban sensing. In 2020 17th Annual IEEE international conference on sensing, communication, and networking (SECON), pages 1–9. I...

    2020

  15. [23]

    Heterogeneous graph neural networks with post-hoc explanations for multi-modal and explainable land use in- ference, 2024

    Xuehao Zhai, Junqi Jiang, Adam Dejl, Antonio Rago, Fangce Guo, Francesca Toni, and Aruna Sivakumar. Heterogeneous graph neural networks with post-hoc explanations for multi-modal and explainable land use in- ference, 2024

    2024

  16. [24]

    Khulbe, A

    D. Khulbe, A. Belyi, O. Mikeš, and S. Sobolevsky. Mobility networks as a predictor of socioeconomic status in urban systems. In Computational Science and Its Applications – ICCSA 2023, pages 139–157, 2023

    2023

  17. [25]

    Longitudinal employer-household dynamics

    Center for Economic Studies US Census Bureau. Longitudinal employer-household dynamics. https://lehd. ces.census.gov/, 2023. Retrieved August 9, 2023

    2023

  18. [26]

    American community survey data

    US Census Bureau. American community survey data. https://www.census.gov/programs-surveys/ acs/data.html, 2023. Retrieved August 9, 2023

    2023

  19. [27]

    Nyc 311 data

    NYC Open Data. Nyc 311 data. https://data.cityofnewyork.us/Social-Services/NYC-311-Data/ jrb2-thup, 2023. Retrieved August 9, 2023

    2023

  20. [28]

    Direction aware positional and structural encoding for directed graph neural networks

    Yonas Sium, Georgios Kollias, Tsuyoshi Idé, Payel Das, Naoki Abe, Aurélie Lozano, and Qi Li. Direction aware positional and structural encoding for directed graph neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023

    2023

  21. [29]

    Dwivedi and X

    V .P. Dwivedi and X. Bresson. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699, 2020

    2012 arXiv

  22. [30]

    Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022. 16 A PREPRINT - JULY 17, 2025

    2022

  23. [31]

    Rethinking graph transformers with spectral attention

    Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems , 34:21618–21629, 2021

    2021

  24. [32]

    Do transformers really perform badly for graph representation? In Advances in Neural Information Processing Systems, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? In Advances in Neural Information Processing Systems, 2021

    2021

  25. [33]

    Distance encoding: Design provably more powerful neural networks for graph representation learning

    Pan Li, Yanbang Wang, Hongwei Wang, and Jure Leskovec. Distance encoding: Design provably more powerful neural networks for graph representation learning. In Advances in Neural Information Processing Systems , volume 33, page 4465–4478, 2020

    2020

  26. [34]

    Graph neu- ral networks with learnable structural and positional representations

    Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Graph neu- ral networks with learnable structural and positional representations. In International Conference on Learning Representations, 2022

    2022

  27. [35]

    Rewiring with positional encodings for graph neural networks

    Rickard Brüel-Gabrielsson, Mikhail Yurochkin, and Justin Solomon. Rewiring with positional encodings for graph neural networks. arXiv preprint arXiv:2201.12674, 2022

    2022 arXiv

  28. [36]

    Sume: Semantic-enhanced urban mobility network embedding for user demographic inference

    Fengli Xu, Zongyu Lin, Tong Xia, Diansheng Guo, and Yong Li. Sume: Semantic-enhanced urban mobility network embedding for user demographic inference. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., 4(3):98, 2020

    2020

  29. [37]

    D. K. Chandra, J. Leopold, and Y . Fu. Nodesense2vec: Spatiotemporal context-aware network embedding for heterogeneous urban mobility data. In 2021 IEEE International Conference on Big Data (Big Data) , pages 2884–2893, 2021

    2021

  30. [38]

    Jain and Richard C

    Anil K. Jain and Richard C. Dubes. Algorithms for clustering data. Prentice-Hall, Inc., 1988

    1988

  31. [39]

    L. Dong, F. Duarte, G. Duranton, and et al. Defining a city — delineating urban areas using cell-phone data. Nature Cities, 1:117–125, 2024

    2024

  32. [40]

    Distance deterrence comparison in urban commute among different socioeconomic groups: A normalized linear piece-wise gravity model

    Mingyi He, Yuri Bogomolov, Devashish Khulbe, and Stanislav Sobolevsky. Distance deterrence comparison in urban commute among different socioeconomic groups: A normalized linear piece-wise gravity model. Journal of Transport Geography, 113:103732, 2023

    2023

  33. [41]

    Impact of income on urban commute across major cities in us

    Yuri Bogomolov, Mingyi He, Devashish Khulbe, and Stanislav Sobolevsky. Impact of income on urban commute across major cities in us. Procedia Computer Science, 193:325–332, 2021

    2021

  34. [42]

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

    2018 arXiv

  35. [43]

    Hicks and P

    N. Hicks and P. Streeten. Indicators of development: The search for a basic needs yardstick. World development, 7(6):567–580, 1979

    1979

  36. [44]

    Donald T. Chen. The science of smart growth. Scientific American, 283(6):84–91, 2000

    2000

  37. [45]

    Leslie and Breandán Ó hUallacháin

    Timothy F. Leslie and Breandán Ó hUallacháin. Polycentric phoenix. Economic Geography, 82(2):167–192, 2006

    2006

  38. [46]

    Structural deep network embedding

    Daixin Wang, Peng Cui, and Wenwu Zhu. Structural deep network embedding. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16) , page 1225–1234, 2016

    2016

  39. [47]

    Aggarwal, and Thomas S

    Shiyu Chang, Wei Han, Jiliang Tang, Guo-Jun Qi, Charu C. Aggarwal, and Thomas S. Huang. Heterogeneous network embedding via deep architectures. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’15), page 119–128, 2015

    2015

  40. [48]

    W. Yap, R. Stouffs, and F. Biljecki. Urbanity: automated modelling and analysis of multidimensional networks in cities. npj Urban Sustainability, 3:45, 2023

    2023

  41. [49]

    Yap and F

    W. Yap and F. Biljecki. A global feature-rich network dataset of cities and dashboard for comprehensive urban analyses. Scientific Data, 10:667, 2023

    2023

  42. [50]

    https://data.boston.gov/dataset/311-service-requests/resource/ f53ebccd-bc61-49f9-83db-625f209c95f5 , 2023

    Boston 311 data. https://data.boston.gov/dataset/311-service-requests/resource/ f53ebccd-bc61-49f9-83db-625f209c95f5 , 2023. Retrieved August 9, 2023

    2023

  43. [51]

    https://data.cityofchicago.org/Service-Requests/311-Service-Requests/ v6vf-nfxy, 2023

    Chicago 311 data. https://data.cityofchicago.org/Service-Requests/311-Service-Requests/ v6vf-nfxy, 2023. Retrieved August 9, 2023

    2023

  44. [52]

    The pagerank citation ranking: bringing order to the web

    Sergey Brin. The pagerank citation ranking: bringing order to the web. Proceedings of ASIS, 98:161–172, 1998

    1998

  45. [53]

    Measuring the vibrancy of urban neighborhoods using mobile phone data with an improved pagerank algorithm

    Chen Jia and et al. Measuring the vibrancy of urban neighborhoods using mobile phone data with an improved pagerank algorithm. Transactions in GIS, 23(2):241–258, 2019. 17 A PREPRINT - JULY 17, 2025

    2019

  46. [54]

    Ranking spaces for predicting human movement in an urban environment

    Bin Jiang. Ranking spaces for predicting human movement in an urban environment. International Journal of Geographical Information Science, 23(7):823–837, 2009. 18

    2009

Pith tools

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