Pith. sign in

REVIEW 4 major objections 5 minor 50 references

MotifGPL: Motif-Enhanced Graph Prototype Learning for Deciphering Urban Social Segregation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read MotifGPL claims that local graph motifs reveal urban segregation and that rewiring graphs along those motifs lowers measured segregation.

desk verdict Worth a look for the motif findings, but the reconstruction experiment is circular and the 'mitigation' claim is overstatement. read the letter →

arxiv 2412.18464 v1 pith:UWQTIYZE submitted 2024-12-24 cs.AI cs.SI

Tengfei He , Xiao Zhou This is my paper · ORCID
classification cs.AIcs.SI
keywords urbansocialsegregationgraphprototypelearningnetworkmotifsspatialorigin-destinationGlobalMoran'sIinterpretableneuralnetworksBeijingblocks
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 local connection patterns of city blocks, expressed as network motifs, carry interpretable information about social segregation and can be used to propose structural changes that reduce it. The authors build MotifGPL, which learns graph prototypes from Beijing's spatial adjacency graph and origin-destination mobility graph, projects each prototype onto real local subgraphs, and converts the projections into motif distributions over 3-, 4-, and selected 5-node motifs. They report that high-segregation blocks cluster in circular, enclosed spatial motifs and long chain-like mobility motifs, while low-segregation blocks show chain-like spatial patterns and simpler star-like commuting patterns. Using the motif distributions to rewire the graphs lowers Global Moran's I from 0.4159 to 0.3169 for the spatial graph and from 0.2410 to 0.1805 for the OD graph, which the authors read as actionable guidance for urban planning. The model also outperforms GNN and interpretable-prototype baselines on segregation-level classification, with accuracy 0.7990.

What carries the argument

The load-bearing object is the motif distribution of a prototype: a frequency count of 3-node, 4-node, and selected 5-node motifs in the local subgraph matched to that prototype. It is produced by a pipeline of GNN encoders on the spatial and OD graphs, a prototype layer with cluster, separation, and encoding losses, a random-walk local structure extractor encoded by an RNN, and a statistical motif test against random graphs. The distribution then drives the reconstruction update $A[i] = (1-\alpha\,\mathrm{KL})A[i] + \alpha\,\mathrm{KL}\,A[\mathrm{tar}]$ followed by the threshold $A^{\mathrm{new}}_{ij} = 1$ if $A_{ij} > \beta$ and $0$ otherwise, where $\mathrm{KL}$ is the KL divergence between motif distributions.

What would settle it

Take the reconstructed adjacency matrices and repeat the rewiring with random degree-preserving edge changes of the same scale: if random rewiring achieves the same reduction in Global Moran's I, the motif guidance is not doing causal work; alternatively, recompute a segregation measure that depends on both node attributes and edge weights and check whether the reported drop from 0.4159 to 0.3169 survives.

Watch

Extended reading notes

Core claim

The central claim is that the motif distribution attached to each learned prototype is the structural signature of urban social segregation. In the spatial graph, high-segregation blocks are dominated by the circular motifs M4,4 and M3,2, suggesting enclosed community structures, whereas low-segregation blocks concentrate in the chain-like motifs M4,1 and M3,1. In the OD graph, high-segregation blocks contain more chain-like motifs and more complex star-like motifs, which the authors interpret as longer commutes and a sharper separation between living and working spaces. Feeding these motif distributions into a graph reconstruction step, the model reduces Global Moran's I while changing less than 2.5 percent of edges at the mildest reconstruction level (alpha = 0.8, beta = 0.3). The authors conclude that MotifGPL reveals the key motifs affecting urban social segregation and provides robust guidance for mitigating it.

Load-bearing premise

The reconstruction experiment assumes that rewriting graph edges according to the learned motif distributions reduces real social segregation, but the only evidence is Global Moran's I computed on the rewritten adjacency matrix with the per-block segregation indices held fixed.

Editorial extensions

If this is right

  • If the central claim is right, urban planners can identify segregation-relevant neighborhoods by matching blocks to a small set of interpretable motifs rather than to opaque statistical indexes.
  • The reconstruction results imply that modest connectivity changes, such as adding or reinforcing edges between blocks with complementary motif distributions, can lower Global Moran's I from 0.4159 to 0.3169 in the spatial graph.
  • The motif signatures for high-segregation blocks (circular spatial motifs, long chain-like commuting motifs) give concrete, testable targets for housing placement and transit investment.
  • The framework extends self-explaining prototype methods from graph classification to node-level urban tasks, so the same architecture can be applied to other socioeconomic outcomes measured at block level.
  • Ablation results imply that both spatial proximity and mobility structure are necessary: removing the spatial graph drops accuracy to 0.7212, and removing the OD graph drops it to 0.7776.

Reading between the lines

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

  • Because the reconstruction is evaluated only by Global Moran's I on the same rewritten adjacency matrix while per-block income, education, and age indices are held fixed, much of the reported drop could be a mathematical consequence of reweighting edges toward similar blocks rather than evidence that real segregation would fall; an external test with actual post-intervention data would settle this
  • The distinction between enclosed circular motifs in segregated areas and chain-like motifs in mixed areas resembles longstanding findings on spatial autocorrelation and income clustering, so the novel contribution may lie in the prototype-to-motif pipeline rather than in the discovery of the patterns themselves.
  • A natural extension is to apply the same prototype-to-motif mapping to time-varying or directed OD graphs to see whether motif distributions shift after a real policy intervention such as a new transit line.
  • Cross-city transfer is a testable consequence: if prototype projections remain stable, motif distributions learned in Beijing could be mapped onto other cities, but that requires new data beyond the paper's single-city study.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces MotifGPL, a graph prototype learning framework for analyzing urban social segregation from a spatial graph and an origin-destination graph of Beijing. The framework learns prototype vectors per segregation class, projects them onto local subgraphs, extracts motif distributions, and then uses those distributions to rewire the two adjacency matrices. The authors report improved classification accuracy over GNN baselines (Table 1), descriptive motif differences between high- and low-segregation blocks (Figure 3), and a decrease in Global Moran's I after graph reconstruction (Table 2), which they interpret as evidence that the model offers actionable strategies for reducing segregation.

Significance. The classification results and the qualitative motif analysis are potentially informative for urban computing, and the paper makes its code publicly available. However, the paper's strongest advertised contribution—that motif-guided reconstruction reduces segregation—rests on an experiment that computes Moran's I on the very adjacency matrix the model rewrites, with per-block segregation values fixed. The reported decrease is therefore a near-tautological consequence of the rewiring, not an empirical finding. The motif analysis itself is descriptive and lacks statistical validation. Because the central mitigation claim is not supported, the paper in its current form does not meet the bar for publication.

major comments (4)
  1. [Urban Graph Structure Reconstruction, Eqs (10)-(13), Table 2] The reconstruction experiment evaluates changes in Global Moran's I computed from the same adjacency matrix that Eqs (10)-(13) modify, while the per-block segregation indices (income, education, age) are held fixed. Moran's I measures the autocorrelation of these fixed node attributes with respect to the spatial weights matrix; any rewiring that adds edges between blocks with dissimilar segregation values will mechanically lower the metric. Consequently, the decrease from 0.4159 to 0.3169 in Table 2 does not demonstrate that segregation as experienced by residents has been reduced. A control condition—e.g., random edge rewiring with the same edge counts, or rewiring by a simple dissimilarity heuristic—is required to show that the motif-based choice of edges matters. Without such a baseline, the abstract and Conclusion's claim of 'robust guidance for mitigating' segregation is unsupported.
  2. [Urban Graph Structure Reconstruction, Eqs (10)-(13)] The update rule in Eq (11) uses A[tar] and mG_tar, but the target node 'tar' is never defined. The reader cannot tell which blocks are paired, how the target is selected, or whether the update is applied sequentially or in parallel. Additionally, the coefficient αKL in Eq (11) is a product of a positive weight and a KL divergence, and may exceed 1, possibly making the convex combination invalid; the authors do not state constraints on α or KL. These details are necessary to reproduce the reported AEP/REP/UEP numbers in Table 2.
  3. [Motif Distribution Discovery, Eq (9), Figure 3] Eq (9) defines motifs as substructures that are statistically overrepresented relative to random networks, but the paper never reports applying this significance test. The distributions in Figure 3 appear to be raw motif counts, and the comparisons between high- and low-segregation blocks are made by visual inspection without error bars or statistical tests. The interpretability claim—that specific motifs characterize segregation—requires at least a permutation or z-score analysis to rule out chance differences.
  4. [Urban Graph Reconstruction, Table 2] The reconstruction experiment reports only MotifGPL's results. There is no comparison to alternative rewiring strategies (random, degree-preserving, or based on node attribute similarity), nor any sensitivity analysis for the free parameters α, β, and Nproto. The claim that motif distributions provide 'novel insights' for reconstruction requires showing that the motif-guided rewiring outperforms simpler heuristics at lowering Moran's I under matched edge-change budgets.
minor comments (5)
  1. [Table 1] The header misspells 'Segregation' as 'Segragation'.
  2. [Eq (1)] The subscript i is missing from τci, and the symbol c is used both for the dimension and as the normalization constant; please clarify the notation.
  3. [Problem Statement] dSEG is referred to as the degree of social segregation but is not formally defined; it should be linked to S_i from Eq (1).
  4. [Eq (7)] The edge weight ωj is introduced but the text does not specify how it is computed or normalized in the random walk.
  5. [Motif Distribution Discovery] The random-walk-based local structure extractor produces T_i ∈ R^{r×t}, but the RNN encoder's input/output dimensions are not specified, making the projection in Eq (5) difficult to reproduce.

Circularity Check

1 steps flagged · score 7.0 of 10

The mitigation claim is circular: Global Moran's I is recomputed on the same adjacency matrix that Eqs. 10-13 rewrite, with per-block segregation indices fixed, so the reported decrease is a mathematical consequence of changing spatial weights rather than evidence of reduced segregation.

  1. self definitional [Section 'Urban Graph Structure Reconstruction', Eqs. (10)-(13); Section 'Experiments', 'Urban Graph Reconstruction', Table 2.]
    "With a reconstruction threshold α, we incrementally adjust the adjacency matrix A ... A[i] = (1− αKL)A[i] + (αKL)A[tar] ... Anew ij = 1 if Aij > β, 0 if Aij ≤ β. ... we utilize Global Moran’s I to measure the overall degree of social segregation across Beijing ... The results indicate that using motif distribution to guide spatial or OD graph reconstruction reduces social segregation."

    Global Moran's I is a weighted spatial autocorrelation of per-block segregation indices, and those indices (computed from socioeconomic indicators via Eq. 1) are never changed during reconstruction. Eqs. 10-13 modify only the adjacency matrix, and the evaluation then computes Moran's I on that same modified matrix. Rewiring to connect blocks with different segregation values necessarily changes the weighted autocorrelation; the reported drop (e.g., 0.4159 to 0.3169 in Gs) is an arithmetic consequence of the new weights, not an observed change in residents' segregation. No random-rewiring or dissimilarity-only baseline is reported, so the motif guidance is not isolated as the cause. The mitigation experiment thus measures the metric on the output of the operation the model itself performs.

full rationale

The classification result (Table 1) and the motif-distribution descriptions are not circular: the prototype model is trained on segregation labels and evaluated on held-out nodes, and the motif analysis is an interpretability exercise on the trained prototypes. The circularity is concentrated in the 'Urban Graph Structure Reconstruction' experiment, which supports the paper's headline contribution of offering 'robust guidance for mitigating' segregation. There, the outcome variable (Global Moran's I) is computed on the adjacency matrix that the reconstruction module itself rewrites. Since the per-block segregation indices are fixed, the drop in Moran's I reflects modified spatial weights, not changed socioeconomic conditions or mobility experiences. Without a control that adds the same number of edges randomly or by a generic dissimilarity heuristic, the claim that motif patterns causally guide mitigation is not supported by this experiment. Self-citations (e.g., He et al. 2020 and Zhou's prior work) appear only in related work and are not load-bearing. The reconstruction evaluation is the one load-bearing circular step; hence a score of 7 rather than 0 or 2.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The framework mostly reuses standard GNN and prototype-learning machinery. The truly new load is the assumption that motif-guided rewiring of the adjacency matrix reduces segregation, which is also where the paper's evidence is weakest. Several hyperparameters (Nproto, alpha, beta, loss weights, walk length) are chosen by hand without sensitivity analysis.

free parameters (7)
  • Nproto (prototypes per class) = 5
    Set based on preliminary experiments to balance interpretability and training time; Section 'Experiment Settings'.
  • alpha (reconstruction weight) = 0.8
    Reconstruction weight factor in Eq 11; Table 2 reports alpha=0.8 with beta varied, no grid search shown.
  • beta (edge creation threshold) = 0.1, 0.2, 0.3
    Threshold in Eq 13; authors report favorable values in Table 2.
  • lambda_1, lambda_2, lambda_3 (loss weights) = 0.4, 0.2, 2
    Hyperparameters of the objective in Eq 18; set without sensitivity analysis.
  • random walk length t and count r = not specified
    Eqs 7-8 define random walk sampling but give no values; needed for the subgraph encoder.
  • spatial graph proximity threshold = not specified
    Spatial edges are built from neighboring areas within a threshold; the threshold value is not reported.
  • segregation label quantile split = median (implied)
    Blocks are categorized into two classes based on quantiles of the segregation index; the specific quantile is not stated.
assumptions (6)
  • standard math GNN message passing correctly embeds graph structure (Eq 3)
    Standard GCN-style propagation treated as given.
  • domain assumption Tobler's first law justifies spatial graph construction from proximity
    Section 'Graph Encoder': spatial graph links neighboring areas within a threshold.
  • domain assumption Segregation index of Moro et al. (2021) is a valid per-block measure
    Eq 1 used as the supervisory signal and to define high/low segregation.
  • domain assumption Network motifs are meaningful units for urban structure interpretation
    Section 'Network Motifs' and 'Motifs Detection' rely on Milo et al. motif concept.
  • ad hoc to paper Rewiring the graph based on motif distributions reduces social segregation
    Section 'Urban Graph Structure Reconstruction', Eqs 10-13; never validated against external outcomes.
  • ad hoc to paper Random-walk subgraph sequences capture the local structure relevant to prototypes
    Eqs 7-8 and 'Prototype Projection': the local structure is represented by RNN-encoded random walks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MotifGPL: Motif-Enhanced Graph Prototype Learning for Deciphering Urban Social Segregation." pith.science (2026). https://pith.science/paper/UWQTIYZE

@misc{pith2026241218464,
  author       = {Pith},
  title        = {Pith review of: MotifGPL: Motif-Enhanced Graph Prototype Learning for Deciphering Urban Social Segregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWQTIYZE}},
  note         = {Machine review of arXiv:2412.18464}
}
read the original abstract

Social segregation in cities, spanning racial, residential, and income dimensions, is becoming more diverse and severe. As urban spaces and social relations grow more complex, residents in metropolitan areas experience varying levels of social segregation. If left unaddressed, this could lead to increased crime rates, heightened social tensions, and other serious issues. Effectively quantifying and analyzing the structures within urban spaces and resident interactions is crucial for addressing segregation. Previous studies have mainly focused on surface-level indicators of urban segregation, lacking comprehensive analyses of urban structure and mobility. This limitation fails to capture the full complexity of segregation. To address this gap, we propose a framework named Motif-Enhanced Graph Prototype Learning (MotifGPL),which consists of three key modules: prototype-based graph structure extraction, motif distribution discovery, and urban graph structure reconstruction. Specifically, we use graph structure prototype learning to extract key prototypes from both the urban spatial graph and the origin-destination graph, incorporating key urban attributes such as points of interest, street view images, and flow indices. To enhance interpretability, the motif distribution discovery module matches each prototype with similar motifs, representing simpler graph structures reflecting local patterns. Finally, we use the motif distribution results to guide the reconstruction of the two graphs. This model enables a detailed exploration of urban spatial structures and resident mobility patterns, helping identify and analyze motif patterns that influence urban segregation, guiding the reconstruction of urban graph structures. Experimental results demonstrate that MotifGPL effectively reveals the key motifs affecting urban social segregation and offer robust guidance for mitigating this issue.

Figures

Figures reproduced from arXiv: 2412.18464 by the authors.

Figure 1
Figure 1. The Framework of Motif-Enhanced Graph Prototype Learning (MotifGPL). [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Network Motifs Employed in the Study.(Blue rep [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Motif Distribution of Prototypes in Gs and Go. Experimental Results Social Segregation Levels Classification Task We assess the model’s performance in predicting segregation levels. The results in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of Reconstruction Outcomes in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 6
Figure 6. Figure 6: Street view selection points within each block. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Examples of street view images in certain blocks. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Motif global distribution map and local spatial [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 39 canonical work pages

  1. [1]

    Baldassarre, F.; and Azizpour, H. 2019. Explainability techniques for graph convolutional networks. arXiv preprint arXiv:1905.13686

    arXiv 2019

  2. [2]

    E.; and Tarabanis, K

    Brimos, P.; Karamanou, A.; Kalampokis, E.; Mamalis, M. E.; and Tarabanis, K. 2023. Explainable Graph Neural Networks on Linked Statistical Data for Predicting Scottish House Prices. In Proceedings of the 27th Pan-Hellenic Conference on Progress in Computing and Informatics, 36--41

    work page 2023

  3. [3]

    A.; and Chung, S.-Y

    Brown, L. A.; and Chung, S.-Y. 2006. Spatial segregation, segregation indices and the geographical perspective. Population, space and place, 12(2): 125--143

    work page 2006

  4. [4]

    Bursell, M.; and Bygren, M. 2023. The making of ethnic segregation in the labor market: Evidence from a field experiment. Technical report, Working Paper

    work page 2023

  5. [5]

    W.; Cai, H.; Chang, K

    Cavallari, S.; Zheng, V. W.; Cai, H.; Chang, K. C.-C.; and Cambria, E. 2017. Learning community embedding with community detection and node embedding on graphs. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, 377--386

    work page 2017

  6. [6]

    Y.; Wang, Y.; Wang, D.; Li, Q.; Lam, W

    Chen, B. Y.; Wang, Y.; Wang, D.; Li, Q.; Lam, W. H.; and Shaw, S.-L. 2018. Understanding the impacts of human mobility on accessibility using massive mobile phone tracking data. Annals of the American Association of Geographers, 108(4): 1115--1133

    work page 2018

  7. [7]

    Chen, J.; and Ying, R. 2024. Tempme: Towards the explainability of temporal graph neural networks via motif discovery. Advances in Neural Information Processing Systems, 36

    work page 2024

  8. [8]

    Ding, D.; Zheng, Y.; Zhang, Y.; and Liu, Y. 2024. Understanding attractions’ connection patterns based on intra-destination tourist mobility: A network motif approach. Humanities and Social Sciences Communications, 11(1): 1--12

    work page 2024

Show all 50 references
  1. [9]

    Y.; and Mostafavi, A

    Fan, C.; Xu, J.; Natarajan, B. Y.; and Mostafavi, A. 2023. Interpretable machine learning learns complex interactions of urban features to understand socio-economic inequality. Computer-Aided Civil and Infrastructure Engineering, 38(14): 2013--2029

    2023

  2. [10]

    Florida, R. 2017. The new urban crisis: How our cities are increasing inequality, deepening segregation, and failing the middle class-and what we can do about it. Hachette UK

    2017

  3. [11]

    Garreton, M.; Basauri, A.; and Valenzuela, L. 2020. Exploring the correlation between city size and residential segregation: comparing Chilean cities with spatially unbiased indexes. Environment and urbanization, 32(2): 569--588

    2020

  4. [12]

    Gottdiener, M.; Hohle, R.; and King, C. 2019. The new urban sociology. Routledge

    2019

  5. [13]

    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, 770--778

    2016

  6. [14]

    He, T.; Bao, J.; Li, R.; Ruan, S.; Li, Y.; Song, L.; He, H.; and Zheng, Y. 2020. What is the human mobility in a new city: Transfer mobility knowledge across cities. In Proceedings of The Web Conference 2020, 1355--1365

    2020

  7. [15]

    Huang, L.; Xia, F.; Chen, H.; Hu, B.; Zhou, X.; Li, C.; Jin, Y.; and Xu, Y. 2023 a . Reconstructing human activities via coupling mobile phone data with location-based social networks. Travel behaviour and society, 33: 100606

    2023

  8. [16]

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

    2023

  9. [17]

    Jin, G.; Liang, Y.; Fang, Y.; Shao, Z.; Huang, J.; Zhang, J.; and Zheng, Y. 2023. Spatio-temporal graph neural networks for predictive learning in urban computing: A survey. IEEE Transactions on Knowledge and Data Engineering

    2023

  10. [18]

    Kakkad, J.; Jannu, J.; Sharma, K.; Aggarwal, C.; and Medya, S. 2023. A survey on explainability of graph neural networks. arXiv preprint arXiv:2306.01958

    2023 arXiv

  11. [19]

    Khoshraftar, S.; and An, A. 2024. A survey on graph representation learning methods. ACM Transactions on Intelligent Systems and Technology, 15(1): 1--55

    2024

  12. [20]

    N.; and Welling, M

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

    2016 arXiv

  13. [21]

    Kwan, M.-P. 2013. Beyond space (as we knew it): Toward temporally integrated geographies of segregation, health, and accessibility: Space--time integration in geography and GIScience. Annals of the Association of American Geographers, 103(5): 1078--1086

    2013

  14. [22]

    C.; and Monkkonen, P

    Lens, M. C.; and Monkkonen, P. 2016. Do strict land use regulations make metropolitan areas more segregated by income? Journal of the American Planning Association, 82(1): 6--21

    2016

  15. [23]

    Li, Y.; Huang, W.; Cong, G.; Wang, H.; and Wang, Z. 2023. Urban region representation learning with openstreetmap building footprints. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 1363--1373

    2023

  16. [24]

    Li, Z.; Huang, W.; Zhao, K.; Yang, M.; Gong, Y.; and Chen, M. 2024. Urban Region Embedding via Multi-View Contrastive Prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 8724--8732

    2024

  17. [25]

    S.; and Denton, N

    Massey, D. S.; and Denton, N. A. 1988. The dimensions of residential segregation. Social forces, 67(2): 281--315

    1988

  18. [26]

    Milo, R.; Shen-Orr, S.; Itzkovitz, S.; Kashtan, N.; Chklovskii, D.; and Alon, U. 2002. Network motifs: simple building blocks of complex networks. Science, 298(5594): 824--827

    2002

  19. [27]

    Moro, E.; Calacci, D.; Dong, X.; and Pentland, A. 2021. Mobility patterns are associated with experienced income segregation in large US cities. Nature communications, 12(1): 4633

    2021

  20. [28]

    Oka, M.; and Wong, D. W. 2019. Segregation: a multi-contextual and multi-faceted phenomenon in stratified societies. In Handbook of urban geography, 255--280. Edward Elgar Publishing

    2019

  21. [29]

    Owens, A.; and Rich, P. 2023. Little boxes all the same? Racial-ethnic segregation and educational inequality across the urban-suburban divide. RSF: The Russell Sage Foundation Journal of the Social Sciences, 9(2): 26--54

    2023

  22. [30]

    Quillian, L. 2014. Does segregation create winners and losers? Residential segregation and inequality in educational attainment. Social Problems, 61(3): 402--426

    2014

  23. [31]

    C.; Hagenbuchner, M.; and Monfardini, G

    Scarselli, F.; Gori, M.; Tsoi, A. C.; Hagenbuchner, M.; and Monfardini, G. 2008. The graph neural network model. IEEE transactions on neural networks, 20(1): 61--80

    2008

  24. [32]

    L.; and Winn, R

    Seewaldt, V. L.; and Winn, R. A. 2023. Residential racial and economic segregation and cancer mortality in the US—speaking out on inequality and injustice. JAMA oncology, 9(1): 126--127

    2023

  25. [33]

    Seo, S.; Kim, S.; and Park, C. 2024. Interpretable prototype-based graph information bottleneck. Advances in Neural Information Processing Systems, 36

    2024

  26. [34]

    Sousa, S.; and Nicosia, V. 2022. Quantifying ethnic segregation in cities through random walks. Nature Communications, 13(1): 5809

    2022

  27. [35]

    Tang, J.; Xia, L.; and Huang, C. 2023. Explainable Spatio-Temporal Graph Neural Networks. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 2432--2441

    2023

  28. [36]

    Veli c kovi \'c , P.; Cucurull, G.; Casanova, A.; Romero, A.; Lio, P.; and Bengio, Y. 2017. Graph attention networks. arXiv preprint arXiv:1710.10903

    2017 arXiv

  29. [37]

    E.; Small, M

    Wang, Q.; Phillips, N. E.; Small, M. L.; and Sampson, R. J. 2018. Urban mobility and neighborhood isolation in America’s 50 largest cities. Proceedings of the National Academy of Sciences, 115(30): 7735--7740

    2018

  30. [38]

    S.; and Williams, D

    White, K.; Haas, J. S.; and Williams, D. R. 2012. Elucidating the role of place in health care disparities: the example of racial/ethnic residential segregation. Health services research, 47(3pt2): 1278--1299

    2012

  31. [39]

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

    2018 arXiv

  32. [40]

    Xu, Z.; and Zhou, X. 2024. CGAP: Urban Region Representation Learning with Coarsened Graph Attention Pooling. arXiv preprint arXiv:2407.02074

    2024 arXiv

  33. [41]

    Yabe, T.; Bueno, B. G. B.; Dong, X.; Pentland, A.; and Moro, E. 2023. Behavioral changes during the COVID-19 pandemic decreased income diversity of urban encounters. Nature communications, 14(1): 2310

    2023

  34. [42]

    Yan, Y.; Wen, H.; Zhong, S.; Chen, W.; Chen, H.; Wen, Q.; Zimmermann, R.; and Liang, Y. 2024. Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web. In Proceedings of the ACM on Web Conference 2024, 4006--4017

    2024

  35. [43]

    Ying, Z.; Bourgeois, D.; You, J.; Zitnik, M.; and Leskovec, J. 2019. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems, 32

    2019

  36. [44]

    Yong, X.; and Zhou, X. 2024. MuseCL: Predicting Urban Socioeconomic Indicators via Multi-Semantic Contrastive Learning. arXiv preprint arXiv:2407.09523

    2024

  37. [45]

    Zhang, S.; Liu, Y.; Shah, N.; and Sun, Y. 2022 a . Gstarx: Explaining graph neural networks with structure-aware cooperative games. Advances in Neural Information Processing Systems, 35: 19810--19823

    2022

  38. [46]

    W.; and Lu, Y

    Zhang, T.; Duan, X.; Wong, D. W.; and Lu, Y. 2021. Discovering income-economic segregation patterns: A residential-mobility embedding approach. Computers, Environment and Urban Systems, 90: 101709

    2021

  39. [47]

    Zhang, Y.; Guo, K.; and Zhou, X. 2024. Causally Aware Generative Adversarial Networks for Light Pollution Control. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 22529--22537

    2024

  40. [48]

    Zhang, Z.; Liu, Q.; Wang, H.; Lu, C.; and Lee, C. 2022 b . Protgnn: Towards self-explaining graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 9127--9135

    2022

  41. [49]

    Zhou, Q.; Lu, X.; Gu, J.; Zheng, Z.; Jin, B.; and Zhou, J. 2024. Explainable Origin-Destination Crowd Flow Interpolation via Variational Multi-Modal Recurrent Graph Auto-Encoder. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 9422--9430

    2024

  42. [50]

    Zou, X.; Yan, Y.; Hao, X.; Hu, Y.; Wen, H.; Liu, E.; Zhang, J.; Li, Y.; Li, T.; Zheng, Y.; et al. 2024. Deep learning for cross-domain data fusion in urban computing: Taxonomy, advances, and outlook. arXiv preprint arXiv:2402.19348

    2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.