Pith. sign in

REVIEW 5 major objections 8 minor 73 references

GeoAI-Enhanced Community Detection on Spatial Networks with Graph Deep Learning

T0 review · 5 major / 8 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper argues that regionalization should pair travel flows with demographic similarity, and that its region2vec embedding method, especially the flow-weighted attention version, outperforms existing community detection baselines and…

desk verdict A solid extension of region2vec to GAT variants, but the headline 'best' result is partly built into the loss, so the paper needs an independent evaluation before that claim can stand. read the letter →

arxiv 2411.15428 v1 pith:WFHEFX2E submitted 2024-11-23 cs.SI cs.AI

Yunlei Liang , Jiawei Zhu , Wen Ye , Song Gao This is my paper · ORCID
classification cs.SIcs.AI
keywords GeoAIcommunitydetectionspatialnetworksgraphattentionconvolutionalnodeembeddingsregionalizationhealthserviceareas
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 regional communities in a spatial network should be defined by travel flows and demographic similarity at once, and that graph neural networks can learn embeddings that honor both. It introduces region2vec, an unsupervised method that trains GCN and GAT encoders with a custom loss: census tracts connected by strong flows are pulled together in embedding space, tract pairs with zero flows are pushed apart, and tracts farther than a hop threshold are penalized for being similar. Geographic adjacency enters through the graph structure and through a connectivity constraint during bottom-up clustering. On Wisconsin census-tract data, the weighted-GAT variant reaches the best combined score, and when used to delineate Health Professional Shortage Areas it flags substantially more shortage area than flow-only baselines. The point is that regionalization can be automated in a repeatable way while preserving both interaction intensity and attribute homogeneity.

What carries the argument

The load-bearing mechanism is the community-oriented loss function combined with graph attention as the encoder. The loss has two terms: a flow-driven term that brings nodes with positive spatial interaction close and pushes zero-flow pairs apart, and a hop-distance term that penalizes embeddings of nodes whose shortest-path distance exceeds a threshold. The weighted GAT variant, a Graph Attention Network whose attention coefficients are multiplied by normalized flow weights, gives intense interaction edges a larger influence on the node embedding. Bottom-up clustering with a geographic-adjacency connectivity constraint then turns the embeddings into contiguous regions.

What would settle it

Run region2vec with weighted GAT on a synthetic spatial network with known planted communities and check whether the recovered partition matches the planted one; a mismatch would show that the community-oriented loss does not reliably balance flows and attributes. Alternatively, compare the Wisconsin communities against an independent flow source such as Medicare patient flows or Census commuting flows, or against the state's own designated service areas; large disagreement would indicate that the SafeGraph-derived pairs are driving the result.

Watch

Extended reading notes

Core claim

The central claim is that an unsupervised, community-oriented embedding loss can make graph neural networks competitive regionalization tools, and that the specific version using attention weighted by spatial flow intensity outperforms traditional community detection and embedding baselines when both flow strength and attribute similarity are wanted. The paper's Table 1 reports an intra-flow ratio of 0.843 and a synthetic score of 0.821 for region2vec with weighted GAT, above Louvain and Leiden (0.801) and all other baselines. As a direct corollary, the authors argue the same pipeline can replace manual Rational Service Area delineation: their method finds nine HPSAs covering 89,848 square kilometers, exceeding the area coverage of other methods while keeping a comparable population-to-provider ratio.

Load-bearing premise

The SafeGraph smartphone visit flows, aggregated to census tracts, are treated as the true spatial interactions that should define regions and health service areas; if those flows are noisy, missing, or biased toward certain business types, the training pairs built from them—pairs with flow used as attractors and zero-flow pairs used as repellers—inherit that bias and the detected communities shift.

Editorial extensions

If this is right

  • Any regionalization task with interaction edges and node attributes can use the same recipe: choose attributes, choose a flow network, train with the community loss, and cluster with a contiguity constraint.
  • Delineation of Rational Service Areas and Health Professional Shortage Areas becomes repeatable and adaptive, since new attributes or new flow networks can be inserted without changing the algorithm.
  • Flow-only modularity methods such as Louvain and Leiden can be beaten on the joint objective, so preserving attribute similarity does not have to come at the expense of interaction intensity.
  • Long-distance mobility flows get reflected in community shape, capturing functional regions such as commuting areas that pure adjacency-based methods miss.

Reading between the lines

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

  • A natural next test is to compare region2vec communities against official state health service areas or against a second flow source such as Medicare claims or commuting data; close agreement would strengthen the case beyond the single-dataset comparison shown here.
  • The same loss could transfer to traffic analysis zone delineation or political redistricting, with the hop-distance term acting as a spatial contiguity regularizer, but those applications are not tested in this paper.
  • Because the loss treats zero-flow pairs as negatives, rural tracts with sparse or missing flow records may be unfairly pushed apart; weighting the negative term by confidence in the zero-flow reading could correct that bias.
  • A synthetic planted-partition experiment could separate the quality of the embedding method from the quality of the flow data by recovering known ground-truth communities exactly.
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

5 major / 8 minor

Summary. The paper proposes a family of unsupervised graph-embedding methods, collectively called region2vec, for community detection on spatial networks. The methods combine node attributes, geographic adjacency, and spatial interaction flows through GCN and GAT encoders, train with a custom loss that pulls flow-positive node pairs together in the embedding space, and produce communities via agglomerative clustering with a spatial-contiguity constraint. The authors compare region2vec variants against Louvain, Leiden, DeepWalk, Node2vec, LINE, and K-Means on SafeGraph-derived flow data for Wisconsin census tracts, reporting that region2vec with weighted GAT achieves the highest intra-flow ratio and highest synthetic score. They also apply the method to delineate Health Professional Shortage Areas and argue that it can support rational service area development. The central claim is that region2vec-weightedGAT simultaneously maximizes node-attribute similarity and spatial-interaction intensity within detected communities.

Significance. If the central claim were supported by an independent evaluation, the paper would make a useful contribution to GeoAI-based regionalization by demonstrating a principled way to integrate multiple graph relationships and node attributes in an unsupervised community-detection pipeline. The authors provide an open-source code repository and include a sensitivity analysis over community counts, which are positive elements. However, the evaluation design currently has a circularity problem: the primary metric, the intra-flow ratio, largely re-measures the quantity minimized by the training loss, and the synthetic score multiplies that metric by attribute-based metrics. As a result, the headline ranking in Table 1 is not an independent test of the claimed advantage. The HPSA case study also lacks external validation. If these issues are addressed, the methodological framework could still be valuable, but the present version does not substantiate the strongest claims in the abstract and conclusion.

major comments (5)
  1. [Section 3.2.1, Eq. (7); Section 3.4.1, Eq. (8); Table 1] The intra-flow ratio (Eq. 8) is effectively the normalized sum of within-community flows, while the first term of the training loss (Eq. 7) minimizes the log-flow-weighted embedding distances of positive-pair nodes (sij > 0). Because the agglomerative clustering in Stage 2 operates on embeddings trained to pull exactly those pairs together, the high intra-flow ratios reported for the region2vec variants in Table 1 are substantially a restatement of the training objective, not an independent measurement of community quality. Louvain and Leiden optimize modularity, a different objective, so the comparison is not apples-to-apples. I request an evaluation that is not directly aligned with the training loss — for example, evaluation on held-out flow data not used in training, comparison against ground-truth or externally defined regions, or a baseline trained with the same loss but without graph-based feature aggregation — or a substantial tempering of the claim that the method is 'best' on the flow dimension.
  2. [Section 5.1.1, Table 1; Section 5.1.2] No repeated-run variability is reported for any method. The differences in intra-flow ratio between region2vec-weightedGAT (0.843) and Louvain (0.829) or Leiden (0.831) may be within random seed variation given the stochastic GNN training. The sensitivity analysis in Section 5.1.2 varies the number of communities but does not vary random seeds, initialization, or model training iterations, so it does not establish stability of the ranking. Please report means and standard deviations over multiple independent runs and, where possible, a statistical test of whether the reported differences are significant.
  3. [Section 4.2, Section 5.2, Table 3] The HPSA case study uses process outputs — number of HPSAs, covered area, and population-to-provider ratio — as evaluation metrics, but these do not validate whether the delineated shortage areas correspond to actual health needs. The statement in Section 5.2 that the shortage has been 'confirmed through discussions with the state health officials' is not a documented external validation. Without comparing the delineated areas to official HRSA-designated HPSAs or to independent health-outcome indicators, the case study demonstrates only that the method can generate such delineations, not that it identifies the correct areas. I recommend adding an external benchmark or clearly relabeling the case study as a feasibility demonstration rather than a validation.
  4. [Section 3.2.1, Algorithm 2; Section 5] The threshold choices t=5 for Spt in the GAT model, t=200 for Spt and t'=100 for S' in the weighted GAT model, the hop-distance threshold epsilon, and the community count K are introduced without justification, and Section 5.1.2 varies only K. These thresholds directly determine the positive-pair set and the adjacency matrix used in Algorithm 2, so the reported performance may be sensitive to them. Please report results across a range of these thresholds or provide a principled justification for the selected values, including how the choice affects the intra-flow ratio and the synthetic score.
  5. [Section 3.2.1, Eq. (7); Table 1; Abstract] The loss function in Eq. (7) contains a spatial-interaction flow term and a hop-distance term, but no explicit attribute-similarity term; node attributes enter only as input features to the GNN. Table 1 shows that K-Means outperforms all region2vec variants on both attribute-only metrics (cosine similarity 0.983 vs. 0.975, and normalized inequality 0 vs. 1.879e-07). The abstract's claim that the proposed method can 'maximize node attribute similarity and spatial interaction intensity simultaneously' is therefore not supported by the training objective or by the attribute-only metrics. Either add an attribute-similarity component to the loss and evaluate its effect, or restrict the claim to 'maintains attribute homogeneity while improving intra-flow performance.'
minor comments (8)
  1. [Section 5.2] The text contains unresolved references 'Figure ??, and Figure ??' in the HPSA results paragraph; these placeholders should be replaced with the actual figure numbers or the figures should be removed.
  2. [Algorithm 1] In Algorithm 1, the denominator of the loss expression uses 'dposq' but Eq. (7) uses 'dnegq' for the negative-pair distances; this typo should be corrected to avoid confusion about the loss definition.
  3. [Section 3.4.4] The synthetic score is described as the product of 'the three metrics,' but the paper defines four individual metrics plus the join count ratio; please clarify why the join count ratio (Eq. 13) is excluded from the synthetic score and whether its inclusion would change the ranking.
  4. [Section 3.4.2, Eq. (9)] Equation (9) appears to assume normalized binary-valued features for the inequality calculation, but the attribute matrix X is not defined as binary; please clarify how non-binary attributes are transformed before computing sigma and mu.
  5. [Section 5.1.2, Figure 3 caption] The caption refers to a '95% confidence-level standard deviation ellipse,' which conflates a confidence ellipse with a standard-deviation ellipse; please clarify what the ellipse represents and how it is computed.
  6. [Section 5.2, Table 3 caption] The caption states 'The HPSA delineation performance of four methods,' but Table 3 lists five methods; either correct the count or restructure the table.
  7. [Section 2, paragraph 1] The sentence 'we first summarize the related literature on communication detection and graph embedding' should read 'community detection' rather than 'communication detection.'
  8. [Figure 1 caption] The caption uses 'regions2vec' but the method name used throughout the paper is 'region2vec'; please make the naming consistent.

Circularity Check

2 steps flagged · score 6.0 of 10

The intra-flow superiority claim re-measures the training objective: Eq. 7 pulls positive-flow pairs together and Eq. 8 scores exactly that, so the headline 'best' result is partially built into the loss.

  1. fitted input called prediction [Section 3.2.1 (Eq. 7) and Section 3.4.1 (Eq. 8)]
    "The spatial interaction flow constraint will draw nodes with spatial interactions (positive pairs) closer and push nodes without spatial interactions (negative pairs) further in the embedding space. ... L = (1/Npos) sum log(sp) dpos_p / ((1/Nneg) sum dneg_q + Lhops). ... RIntraFlow = sum_{ci=cj} s_ij / sum s_ij; ci, cj in {1,2,...,K}."

    By Eq. 7, the training loss minimizes Euclidean embedding distances between positive flow pairs, weighted by log(s_p), and maximizes negative-pair distances. Eq. 8 then scores any partition by the fraction of total flow weight internal to communities. Since agglomerative clustering partitions exactly the embedding space in which those positive pairs were pulled together, a high intra-flow ratio for region2vec is the training objective re-expressed as an evaluation metric, not an independent confirmation. The comparison is also asymmetric: Louvain and Leiden optimize modularity, a different objective, while region2vec is explicitly trained to maximize the quantity that R_IntraFlow measures.

  2. self definitional [Section 3.4.4 (Eq. 12) and Section 5.1.1 (Table 1)]
    "S = RIntraFlow * CosineSimilarity * (1 - Inorm). ... Overall, the proposed region2vec method with weighted GAT has the best performance for the intra-flow ratio and the synthetic score and the second-best performance for inequality and cosine similarity."

    The synthetic score multiplies the already-circular intra-flow term by attribute metrics, so the 'best overall synthetic score' inherits the training objective as a major factor. At the same time, Eq. 7 contains no explicit attribute-similarity loss; node attributes enter only as GNN input features. Table 1 shows K-Means actually beats region2vec on both attribute metrics (cosine similarity 0.983 vs 0.975; normalized inequality 0 vs 1.879E-07). Therefore the headline conclusion of simultaneously maximizing attribute similarity and spatial interaction intensity is not independently demonstrated: the flow component is essentially defined by the training loss, and the attribute component is not explicitly optimized by that loss.

full rationale

The paper is not wholly circular: the GNN embeddings do encode node attributes through input features, the attribute metrics are not directly optimized in Eq. 7, and the connectivity-constrained agglomerative clustering adds a nontrivial post-processing step. However, the central performance claim is substantially undermined by metric-loss alignment. Eq. 7 directly minimizes flow-weighted embedding distances for positive flow pairs, while Eq. 8 measures exactly the share of flow weight internal to the detected communities, so the intra-flow score for region2vec is largely a restatement of the training objective. The synthetic score (Eq. 12) imports this same term, making the 'best overall' conclusion partly built into the evaluation. The attribute component does provide independent evidence, and K-Means outperforms region2vec on both attribute metrics, which shows the joint claim is only partially forced. Self-citations to Liang et al. (2022) for the GCN model and the earlier intra-inter flow metric are transparent and are not the main load-bearing circularity. Score 6 reflects partial circularity where one key prediction reduces by construction, rather than a fully self-referential derivation.

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

The central comparison depends on several user-chosen thresholds and hyperparameters: t (flow threshold), t' (normalization threshold), epsilon (hop penalty), K (community count), and the network architecture. None are learned from data or justified by an optimization procedure in the paper, and the key evaluation metric overlaps with the loss objective. The SafeGraph flow assumption is an unvalidated domain assumption. No new entities are introduced.

free parameters (6)
  • Flow threshold t for the GAT adjacency matrix (Spt) = 5
    Section 5.1: 'threshold t in Spt is set to 5 to include most of the positive flow edges'; value affects which flow edges enter AGAT in Eq. 6; no sensitivity analysis reported.
  • Flow threshold t for weighted GAT (Spt) = 200
    Section 5.1: set to 200; much larger than the GAT threshold; no tuning procedure or sensitivity analysis.
  • Normalized flow threshold t' for S' in weighted GAT = 100
    Section 5.1: t' set to 100; determines which flow weights multiply attention coefficients in Eq. 5; chosen ad hoc.
  • Hop-distance threshold epsilon in Lhops = not reported
    Equation 7 defines Lhops using I(hop_ij > epsilon); the experimental value of epsilon is not given anywhere in the paper.
  • Community count K = 14 (from Louvain)
    Section 5: K set to 14 for all methods requiring it, matching Louvain's output; sensitivity analysis over K is provided, but K is still an input chosen by the analyst.
  • Graph neural network hyperparameters (hidden size, output dimension, learning rate, epoch count) = not reported
    Section 3.2.1 describes layer counts but provides no hyperparameter table or configuration file; reproducibility of the exact model is limited.
assumptions (5)
  • domain assumption Tobler's First Law of Geography: nearby nodes are more related, so neighborhood aggregation in GCN/GAT is appropriate.
    Invoked in Section 3.2.1 to justify graph convolution for spatial networks.
  • domain assumption SafeGraph business visit flows aggregated to census tracts approximate the spatial interactions relevant to community detection and health service areas.
    Section 4.1 constructs the spatial interaction network from SafeGraph venue visits; no validation of representativeness.
  • domain assumption Pairs with zero flow are true negative interactions and should be pushed apart in embedding space.
    Algorithms 1 and 2 define neg_n as all pairs with s_ik = 0; most census tract pairs have zero flow, making this assumption very strong.
  • ad hoc to paper The product of metrics in Eq. 12 is a valid combined measure of community quality.
    Synthetic score is a new composite defined by the authors; no independent evidence is given that equal weighting is appropriate.
  • standard math Standard GCN and GAT update rules from Kipf and Welling (2017) and Velickovic et al. (2017).
    Section 3.2.1 equations 1-4 rely on the cited models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoAI-Enhanced Community Detection on Spatial Networks with Graph Deep Learning." pith.science (2026). https://pith.science/paper/WFHEFX2E

@misc{pith2026241115428,
  author       = {Pith},
  title        = {Pith review of: GeoAI-Enhanced Community Detection on Spatial Networks with Graph Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFHEFX2E}},
  note         = {Machine review of arXiv:2411.15428}
}
read the original abstract

Spatial networks are useful for modeling geographic phenomena where spatial interaction plays an important role. To analyze the spatial networks and their internal structures, graph-based methods such as community detection have been widely used. Community detection aims to extract strongly connected components from the network and reveal the hidden relationships between nodes, but they usually do not involve the attribute information. To consider edge-based interactions and node attributes together, this study proposed a family of GeoAI-enhanced unsupervised community detection methods called region2vec based on Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN). The region2vec methods generate node neural embeddings based on attribute similarity, geographic adjacency and spatial interactions, and then extract network communities based on node embeddings using agglomerative clustering. The proposed GeoAI-based methods are compared with multiple baselines and perform the best when one wants to maximize node attribute similarity and spatial interaction intensity simultaneously within the spatial network communities. It is further applied in the shortage area delineation problem in public health and demonstrates its promise in regionalization problems.

Figures

Figures reproduced from arXiv: 2411.15428 by the authors.

Figure 1
Figure 1. The workflow for community detection using the regions2vec method. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. The 3D view of three metrics for all the methods. Left: the view of all methods, Right: the zoomed-in [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. The distribution of all methods based on cosine similarity and intra-flow ratio. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The resulting communities maps of all methods. [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: The final Health Professional Shortage Area with scores for three methods. [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 67 canonical work pages

  1. [1]

    and Barab \'a si, A.-L

    Albert, R. and Barab \'a si, A.-L. (2002). Statistical mechanics of complex networks. Reviews of modern physics , 74(1):47

    work page 2002

  2. [2]

    Barth \'e lemy, M. (2011). Spatial networks. Physics reports , 499(1-3):1--101

    work page 2011

  3. [3]

    Batty, M. (2021). Defining urban science. Urban Informatics , pages 15--28

    work page 2021

  4. [4]

    Representation learning in multiplex graphs: Where and how to fuse information?

    Bielak, P. and Kajdanowicz, T. (2024). Representation learning in multiplex graphs: Where and how to fuse information? arXiv preprint arXiv:2402.17906

    work page Pith review arXiv 2024

  5. [5]

    D., Guillaume, J.-L., Lambiotte, R., and Lefebvre, E

    Blondel, V. D., Guillaume, J.-L., Lambiotte, R., and Lefebvre, E. (2008). Fast unfolding of communities in large networks. Journal of statistical mechanics: theory and experiment , 2008(10):P10008

    2008

  6. [6]

    Cliff, A. D. and Ord, J. K. (1973). Spatial autocorrelation . London: Pion

    work page 1973

  7. [7]

    Cui, G., Zhou, J., Yang, C., and Liu, Z. (2020). Adaptive graph encoder for attributed graph embedding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 976--985

    work page 2020

  8. [8]

    J., Sieber, R., Tyukin, I., and Yeboah, G

    De Sabbata, S., Ballatore, A., Miller, H. J., Sieber, R., Tyukin, I., and Yeboah, G. (2023). GeoAI in urban analytics . International Journal of Geographical Information Science , 37(12):2455--2463

    work page 2023

Show all 73 references
  1. [9]

    and Liu, P

    De Sabbata, S. and Liu, P. (2023). A graph neural network framework for spatial geodemographic classification. International Journal of Geographical Information Science , 37(12):2464--2486

    2023

  2. [10]

    S., Blondel, V

    Expert, P., Evans, T. S., Blondel, V. D., and Lambiotte, R. (2011). Uncovering space-independent communities in spatial networks. Proceedings of the National Academy of Sciences , 108(19):7663--7668

    2011

  3. [11]

    W., Tarjan, R

    Flake, G. W., Tarjan, R. E., and Tsioutsiouliklis, K. (2004). Graph clustering and minimum cut trees. Internet Mathematics , 1(4):385--408

    2004

  4. [12]

    Fu, X., Zhang, J., Meng, Z., and King, I. (2020). Magnn: Metapath aggregated graph neural network for heterogeneous graph embedding. In Proceedings of The Web Conference 2020 , pages 2331--2341

    2020

  5. [13]

    Gao, S., Liu, Y., Wang, Y., and Ma, X. (2013). Discovering spatial interaction communities from mobile phone data. Transactions in GIS , 17(3):463--481

    2013

  6. [14]

    Geography for life, national geography standards

    Geography Education Standards Project (1994). Geography for life, national geography standards

    1994

  7. [15]

    and Ferrara, E

    Goyal, P. and Ferrara, E. (2018). Graph embedding techniques, applications, and performance: A survey. Knowledge-Based Systems , 151:78--94

    2018

  8. [16]

    Grekousis, G. (2019). Artificial neural networks and deep learning in urban geography: A systematic review and meta-analysis. Computers, Environment and Urban Systems , 74:244--256

    2019

  9. [17]

    and Leskovec, J

    Grover, A. and Leskovec, J. (2016). node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages 855--864

    2016

  10. [18]

    Hamilton, W., Ying, Z., and Leskovec, J. (2017). Inductive representation learning on large graphs. Advances in neural information processing systems , 30

    2017

  11. [19]

    He, D., Song, Y., Jin, D., Feng, Z., Zhang, B., Yu, Z., and Zhang, W. (2021). Community-Centric Graph Convolutional Network for Unsupervised Community Detection . pages 3515--3521

    2021

  12. [20]

    Shortage Designation Management System (SDMS): Manual for Policies and Procedures

    Health Resources and Services Administration (2020). Shortage Designation Management System (SDMS): Manual for Policies and Procedures . https://contentmanager.med.uvm.edu/docs/sdms_manual_/ahec-documents/sdms_manual_.pdf

    2020

  13. [21]

    Hinton, G. E. and Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. science , 313(5786):504--507

    2006

  14. [22]

    S., and Patz, J

    Hou, X., Gao, S., Li, Q., Kang, Y., Chen, N., Chen, K., Rao, J., Ellenberg, J. S., and Patz, J. A. (2021). Intracounty modeling of covid-19 infection with human mobility: Assessing spatial heterogeneity with business traffic, age, and race. Proceedings of the National Academy ...

    2021

  15. [23]

    The bureau of health workforce portal

    HRSA (2023). The bureau of health workforce portal. https://programportal.hrsa.gov/extranet/landing.seam

    2023

  16. [24]

    Hu, Y., Goodchild, M., Zhu, A.-X., Yuan, M., Aydin, O., Bhaduri, B., Gao, S., Li, W., Lunga, D., and Newsam, S. (2024). A five-year milestone: reflections on advances and limitations in geoai research. Annals of GIS , pages 1--14

    2024

  17. [25]

    Hu, Y., Wang, F., and Xierali, I. M. (2018). Automated delineation of hospital service areas and hospital referral regions by modularity optimization. Health services research , 53(1):236--255

    2018

  18. [26]

    Janowicz, K., Gao, S., McKenzie, G., Hu, Y., and Bhaduri, B. (2020). Geoai: spatially explicit artificial intelligence techniques for geographic knowledge discovery and beyond. International Journal of Geographical Information Science , 34(4):625--636

    2020

  19. [27]

    Jiang, B. (2007). A topological pattern of urban street networks: universality and peculiarity. Physica A: Statistical mechanics and its applications , 384(2):647--655

    2007

  20. [28]

    Jiang, B. (2018). A topological representation for taking cities as a coherent whole. Geographical Analysis , 50(3):335--352

    2018

  21. [29]

    and Ma, D

    Jiang, B. and Ma, D. (2015). Defining least community as a homogeneous group in complex networks. Physica A: Statistical Mechanics and its Applications , 428:154--160

    2015

  22. [30]

    and Ren, Z

    Jiang, B. and Ren, Z. (2019). Geographic space as a living structure for predicting human activities using big data. International Journal of Geographical Information Science , 33(4):764--779

    2019

  23. [31]

    Jin, D., Li, B., Jiao, P., He, D., and Shan, H. (2019). Community Detection via Joint Graph Convolutional Network Embedding in Attribute Network . Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatic...

    2019

  24. [32]

    Johnson, S. C. (1967). Hierarchical clustering schemes. Psychometrika , 32(3):241--254

    1967

  25. [33]

    Kang, Y., Gao, S., Liang, Y., Li, M., Rao, J., and Kruse, J. (2020). Multiscale dynamic human mobility flow dataset in the us during the covid-19 epidemic. Scientific Data , 7(1):1--13

    2020

  26. [34]

    Kang, Y., Wu, K., Gao, S., Ng, I., Rao, J., Ye, S., Zhang, F., and Fei, T. (2022). Sticc: a multivariate spatial clustering method for repeated geographic pattern discovery with consideration of spatial contiguity. International Journal of Geographical Information Science , 36...

    2022

  27. [35]

    Kipf, T. N. and Welling, M. (2017). Semi-supervised classification with graph convolutional networks . 5th International Conference on Learning Representations, ICLR 2017 - Conference Track Proceedings , pages 1--14

    2017

  28. [36]

    P., and Mayer, K

    Kruse, J., Gao, S., Ji, Y., Szabo, D. P., and Mayer, K. R. (2023). Bringing spatial interaction measures into multi-criteria assessment of redistricting plans using interactive web mapping. Cartography and Geographic Information Science , pages 1--20

    2023

  29. [37]

    LeCun, Y., Bengio, Y., et al. (1995). Convolutional networks for images, speech, and time series. The handbook of brain theory and neural networks , 3361(10):1995

    1995

  30. [38]

    Liang, Y., Zhu, J., Ye, W., and Gao, S. (2022). Region2vec: community detection on spatial networks using graph embedding with node attributes and spatial interactions. In Proceedings of the 30th International Conference on Advances in Geographic Information Systems , pages 1--4

    2022

  31. [39]

    Liu, K., Gao, S., and Lu, F. (2019). Identifying spatial interaction patterns of vehicle movements on urban road networks by topic modelling. Computers, Environment and Urban Systems , 74:50--61

    2019

  32. [40]

    and Biljecki, F

    Liu, P. and Biljecki, F. (2022). A review of spatially-explicit geoai applications in urban geography. International Journal of Applied Earth Observation and Geoinformation , 112:102936

    2022

  33. [41]

    Liu, X., Kang, C., Gong, L., and Liu, Y. (2016). Incorporating spatial interaction patterns in classifying and understanding urban land use. International Journal of Geographical Information Science , 30(2):334--350

    2016

  34. [42]

    Liu, Y., Kang, C., Gao, S., Xiao, Y., and Tian, Y. (2012). Understanding intra-urban trip patterns from taxi trajectory data. Journal of geographical systems , 14(4):463--483

    2012

  35. [43]

    Liu, Y., Sui, Z., Kang, C., and Gao, Y. (2014). Uncovering patterns of inter-urban trip and spatial interaction from social media check-in data. PloS One , 9(1):e86026

    2014

  36. [44]

    Lopes, P. M. (2000). State-wide rational service areas for primary care services: Lessons from six states. Technical report

    2000

  37. [45]

    Ma, D., Osaragi, T., Oki, T., and Jiang, B. (2020). Exploring the heterogeneity of human urban movements using geo-tagged tweets. International Journal of Geographical Information Science , 34(12):2475--2496

    2020

  38. [46]

    MacQueen, J. et al. (1967). Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability , volume 1, pages 281--297. Oakland, CA, USA

    1967

  39. [47]

    Mai, G., Janowicz, K., Hu, Y., Gao, S., Yan, B., Zhu, R., Cai, L., and Lao, N. (2022). A review of location encoding for geoai: methods and applications. International Journal of Geographical Information Science , 36(4):639--673

    2022

  40. [48]

    Manning, C., Raghavan, P., and Sch \"u tze, H. (2010). Introduction to information retrieval. Natural Language Engineering , 16(1):100--103

    2010

  41. [49]

    B., Kar, N

    Molokwu, B., Shuvo, S. B., Kar, N. C., and Kobti, Z. (2020). Node classification and link prediction in social graphs using rlvecn. In 32nd International Conference on Scientific and Statistical Database Management , pages 1--10

    2020

  42. [50]

    Nelson, G. D. and Rae, A. (2016). An economic geography of the united states: From commutes to megaregions. PloS One , 11(11):e0166083

    2016

  43. [51]

    Newman, M. E. (2006). Modularity and community structure in networks. Proceedings of the national academy of sciences , 103(23):8577--8582

    2006

  44. [52]

    Pandey, B., Brelsford, C., and Seto, K. C. (2022). Infrastructure inequality is a characteristic of urbanization. Proceedings of the National Academy of Sciences , 119(15):e2119890119

    2022

  45. [53]

    J., Lee, K., and Choi, J

    Park, J., Lee, M., Chang, H. J., Lee, K., and Choi, J. Y. (2019). Symmetric graph convolutional autoencoder for unsupervised graph representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 6519--6528

    2019

  46. [54]

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. (2011). Scikit-learn: Machine learning in P ython. Jou...

    2011

  47. [55]

    Perozzi, B., Al-Rfou, R., and Skiena, S. (2014). Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining , pages 701--710

    2014

  48. [56]

    Pinheiro, D., Hartman, R., Romero, E., Menezes, R., and Cadeiras, M. (2020). Network-based delineation of health service areas: A comparative analysis of community detection algorithms. In Complex Networks XI , pages 359--370. Springer

    2020

  49. [57]

    Ratti, C., Sobolevsky, S., Calabrese, F., Andris, C., Reades, J., Martino, M., Claxton, R., and Strogatz, S. H. (2010). Redrawing the map of great britain from a network of human interactions. PloS one , 5(12):e14248

    2010

  50. [58]

    Power your product with the highest quality pois

    SafeGraph (2023). Power your product with the highest quality pois. https://www.safegraph.com

    2023

  51. [59]

    Su, X., Xue, S., Liu, F., Wu, J., Member, S., Yang, J., Zhou, C., and Hu, W. (2021). A Comprehensive Survey on Community Detection with Deep Learning . (Xx)

    2021

  52. [60]

    Tang, J., Qu, M., Wang, M., Zhang, M., Yan, J., and Mei, Q. (2015). Line: Large-scale information network embedding. In Proceedings of the 24th international conference on world wide web , pages 1067--1077

    2015

  53. [61]

    Tobler, W. R. (1970). A computer movie simulating urban growth in the detroit region. Economic geography , 46(sup1):234--240

    1970

  54. [62]

    A., Waltman, L., and Van Eck, N

    Traag, V. A., Waltman, L., and Van Eck, N. J. (2019). From louvain to leiden: guaranteeing well-connected communities. Scientific reports , 9(1):5233

    2019

  55. [63]

    https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html

    US Census Bureau (2023). Tiger/line shapefiles. "https://www.census.gov/geographies/mapping-files/time-series/geo/tiger-line-file.html"

    2023

  56. [64]

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

    2017

  57. [65]

    and Wang, F

    Wang, C. and Wang, F. (2022). GIS-automated delineation of hospital service areas in Florida: from Dartmouth method to network community detection methods . Annals of GIS , 00

    2022

  58. [66]

    Wang, C., Wang, F., and Onega, T. (2021). Network optimization approach to delineating health care service areas: Spatially constrained Louvain and Leiden algorithms . Transactions in GIS , 25(2):1065--1081

    2021

  59. [67]

    Wang, D., Cui, P., and Zhu, W. (2016). Structural deep network embedding. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining , pages 1225--1234

    2016

  60. [68]

    Xiao, S., Wang, S., Dai, Y., and Guo, W. (2022). Graph neural networks in node classification: survey and evaluation. Machine Vision and Applications , 33:1--19

    2022

  61. [69]

    and Andris, C

    Ye, X. and Andris, C. (2021). Spatial social networks in geographic information science

    2021

  62. [70]

    Zeng, H., Zhou, H., Srivastava, A., Kannan, R., and Prasanna, V. (2019). Accurate, efficient and scalable graph embedding. In 2019 IEEE International Parallel and Distributed Processing Symposium , pages 462--471. IEEE

    2019

  63. [71]

    Zhang, T., Xiong, Y., Zhang, J., Zhang, Y., Jiao, Y., and Zhu, Y. (2020). Commdgi: community detection oriented deep graph infomax. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages 1843--1852

    2020

  64. [72]

    and Turner, M

    Zhu, A.-X. and Turner, M. (2022). How is the third law of geography different? Annals of GIS , 28(1):57--67

    2022

  65. [73]

    Zhu, D., Zhang, F., Wang, S., Wang, Y., Cheng, X., Huang, Z., and Liu, Y. (2020). Understanding place characteristics in geographic contexts through graph convolutional neural networks. Annals of the American Association of Geographers , 110(2):408--420

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.