Pith. sign in

REVIEW 3 major objections 6 minor 42 references

PlaceRep: Geospatial Place Representation Learning from Large-Scale Point-of-Interest Data

T0 review · 3 major / 6 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 PlaceRep claims that region embeddings built by clustering nearby, similarly typed points of interest beat most trained graph-based geospatial baselines on population density and housing price prediction, while running up to 100x faster.

desk verdict PlaceRep's region embedding is a mean of POI vectors regardless of clustering, so its central claim about place-aware representation collapses. read the letter →

arxiv 2507.02921 v4 pith:2AWK7AOJ submitted 2025-06-25 cs.LG cs.AI

Mohammad Hashemi , Hossein Amiri , Andreas Zufle This is my paper
classification cs.LGcs.AI
keywords geospatialrepresentationlearningplacediscoverypoint-of-interestgraphsgraphcondensationtraining-freeembeddingsmulti-granularityurbanregionsimplifiedconvolution
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

PlaceRep's central claim is that meaningful urban places—clusters of nearby, semantically similar points of interest (POIs)—can be discovered without training a model, and that these clusters summarize large POI graphs into region embeddings that are more accurate than those from trained graph neural network baselines. The method propagates POI category features through a spatial graph, clusters the resulting embeddings with bisecting k-means, and averages the cluster centroids to form ZIP-code-level representations. On population density and housing price prediction across U.S. states, the paper reports that PlaceRep outperforms Place2Vec, HGI, PDFM, and simple averaging on most tasks, with reductions in RMSE and gains in $R^{2}$ such as 0.89 versus 0.53 for the best baseline in Georgia. Because the whole pipeline is training-free, the paper reports up to a 100x speedup in generating region representations on large-scale POI graphs, making multi-granular urban analysis feasible at national scale.

What carries the argument

The central object is the 'place'—a cluster of POIs that are both spatially adjacent (via the Delaunay-triangulated graph structure) and semantically similar (via propagated category embeddings). The mechanism that carries the argument is the combination of: (1) a non-parametric SGC propagation that produces neighborhood-aware POI embeddings; (2) bisecting k-means clustering with a reduction ratio $r$ that controls the number and granularity of places per region; and (3) a weighted averaging aggregator that compresses place centroids into a single region embedding. The reduction ratio $r$ is the granularity dial: small values give coarse, broad places, large values give fine-grained, specialized ones. The paper cites a graph-condensation guarantee to assert that these centroids remain as expressive as the original POI embeddings for downstream prediction.

What would settle it

Run a controlled comparison on a single state, e.g. Georgia: generate region embeddings from (a) PlaceRep's cluster centroids, (b) the full set of POI embeddings, and (c) an equal number of clusters formed by randomly partitioning POIs. If the random partitions match or beat PlaceRep's clusters on population density or housing price RMSE, the claim that semantic-spatial places—not mere compression—drive the accuracy is falsified. The same comparison on the cities in Figure 5 would show whether the discovered places correspond to real functional zones.

Watch

Extended reading notes

Core claim

PlaceRep's core discovery is that a place-based, training-free compression of a POI graph carries the information needed for region-level prediction. Starting from SD-CEM hierarchical category features, the method enriches each POI by simplified graph convolution with weighted multi-step propagation, $H = \sum_{k=0}^{K} \alpha_k \hat{A}^k V$, then partitions each administrative region's POI embeddings into $k_r = \lfloor n_r r \rfloor$ clusters via bisecting k-means, and finally aggregates cluster centroids into a region embedding weighted by POI counts. The paper argues this produces not just embeddings but identifiable places—spatially contiguous, semantically coherent units—that are more expressive than fixed administrative aggregates. Empirically, PlaceRep achieves the lowest RMSE and highest $R^{2}$ on most of the seven states tested for both population density and housing price prediction, and it generates Florida's region embeddings more than ten times faster than the compared baselines while also improving accuracy.

Load-bearing premise

The load-bearing premise is that the condensation guarantee proven for graph condensation with GNN training carries over to k-means centroids of SGC-propagated POI embeddings; the paper asserts this transfer of theorem rather than verifying its conditions.

Editorial extensions

If this is right

  • No-training region embeddings make it possible to generate representations for any new region or a national-scale POI graph in minutes on standard CPUs, removing the GPU bottleneck of current pretraining approaches.
  • The reduction ratio $r$ gives a single method that can produce embeddings at any spatial granularity, so one framework can serve tasks at the neighborhood, ZIP-code, or city level without retraining or code changes.
  • Because every POI is explicitly mapped to a place, the resulting representations are traceable and interpretable for urban planning, enabling discovery of functional districts such as mixed dining-and-office corridors.
  • The 100x speedup claim implies that graph-based geospatial representation learning can scale to tens of millions of POIs, a regime where current GNN training on sampled cities is the standard compromise.
  • PlaceRep's strong performance on densely populated states and weak performance on sparse ones (negative R^2 in Wyoming and Vermont) suggests the method's value grows with POI density, pointing to where it should and should not be applied.

Reading between the lines

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

  • Because the expressivity guarantee is taken from graph condensation without verifying its transfer, the accuracy of PlaceRep might come from the SGC smoothing rather than the place clustering; a random-cluster control experiment, which the paper does not run, would separate these.
  • The reported speedups omit the cost of SD-CEM feature pretraining and the CPU clustering time at national scale; a full end-to-end accounting would clarify whether 'training-free' holds once all components are included.
  • The method's limitation on sparse states suggests that a general-purpose geospatial foundation model would need to fuse other modalities, such as satellite imagery or mobility patterns, to reach useful performance where POIs are scarce.
  • If the place clusters do correspond to true functional areas, PlaceRep could serve as a low-cost way to generate training labels or spatial priors for other geospatial models, a transfer the paper does not explore.
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

3 major / 6 minor

Summary. The paper proposes PlaceRep, a training-free pipeline for geospatial region representation learning. POI category vectors from a pretrained SD-CEM encoder are propagated on a Delaunay or KNN POI graph using SGC-style weighted multi-step propagation; bisecting k-means then clusters the propagated embeddings within administrative regions to form 'places,' and a region aggregator produces a single embedding per region. The authors construct FSQ-19M, a ~19M-POI graph over the contiguous U.S., and evaluate the resulting embeddings on population density and housing price prediction against Averaging, Place2Vec, HGI, and PDFM, reporting accuracy gains and large speedups, together with a qualitative place-identification study.

Significance. If supported, the paper would be valuable: a training-free method that matches or beats trained GNNs at lower cost, a large POI graph dataset, and a formal notion of 'place' are all potentially useful contributions. The authors provide code, reimplement baselines, and report 10-run means. However, the algebra of the region aggregator removes the clustering step entirely, so the central claim that place-level clustering improves region representations is not supported by the experiments as described.

major comments (3)
  1. [§4.3.4, Eq. (6)] The region aggregator as written cancels the clustering step. With z_j = (1/|C_j|) Σ_{h∈C_j} h (Eq. 6), the weighted average L_r = Σ_j (|C_j|/Σ_i |C_i|) z_j telescopes to (1/|V_r|) Σ_{h∈V_r} h, the ordinary mean of the propagated POI embeddings. The k-means partition, the reduction ratio r, and the cluster assignments therefore have no effect on the 30-dimensional vectors evaluated in Tables 3, 4, and 6. Any observed gains must be attributed to the SD-CEM encoder and SGC propagation (Eqs. 2–3), not to the proposed place representation learning. The GECC guarantee cited in §4.3.3 does not apply because the condensed set Z_r is averaged into a single vector before any downstream use; no model is trained on Z_r. The alternative reading of L_r = [l1 | l2 | ...] as a concatenation is incompatible with the fixed d=30 embedding dimension stated in §5.1 and with the Random Forest setup.
  2. [Tables 3 and 4] No significance tests are reported, and several PlaceRep advantages are within one standard deviation. For example, Table 3, WY, RMSE is 61.14±38.00 for PlaceRep versus 62.12±32.22 for Place2Vec, and Table 4, NY, PlaceRep RMSE is 418.94±50.77 versus 410.11±49.04 for PDFM, i.e., PlaceRep is worse there. With 10 runs, the claim that PlaceRep 'consistently outperforms' baselines is not established for these states. Please report paired significance tests or confidence intervals and temper the state-level claims accordingly.
  3. [Abstract and §5.3] The abstract claims 'up to a x100 speedup,' but the numbers in §5.3 and Figure 4 do not support this. The largest ratios in the text are about 45.9x (WY: 60.61s / 1.32s for HGI) and about 26.8x (FL: 1293.84s / 48.29s for HGI). No 100x speedup appears. The 'order-of-magnitude' statement in §5.3 is supported, but the abstract's x100 claim should be corrected or substantiated with the specific configuration that yields it.
minor comments (6)
  1. [§4] The word 'gepspatial' in the opening sentence of Section 4 is a typo for 'geospatial.'
  2. [§3.2.2] The phrase 'POIs ro higher-level places' should read 'POIs or higher-level places.'
  3. [§5.2.1] The sentence 'This means that the for these states having too few POIs...' contains a grammatical error; it should be 'This means that for these states, which have too few POIs...'.
  4. [Table 5] In the Wyoming row, the alpha_2 entry is '0.0/0,0,' which contains a typo and an inconsistent separator; it should be a single slash-separated pair like the other entries.
  5. [§4.3.2 and Table 5] The text says alpha_k may be negative to capture heterophily, but the tuning procedure in §5.1.1 and Table 5 restrict alpha to [0.0, 1.0]. The heterophily mechanism is therefore never exercised; either implement negative alpha or remove this claim.
  6. [Abstract and §4.3.1] The abstract says PlaceRep eliminates model pre-training, but the method relies on the pretrained SD-CEM encoder. Recommend rewording to 'no additional training' or explicitly clarifying that the pretrained encoder is external.

Circularity Check

2 steps flagged · score 6.0 of 10

PlaceRep's region aggregator makes clustering algebraically irrelevant: by the paper's own formulas the region embedding is exactly the mean of the propagated POI embeddings, so the central 'place representation learning' step cancels by construction.

  1. self definitional [Section 4.3.3, Eq. (6), and Section 4.3.4 (Region Aggregator)]
    "Each cluster is summarized by its centroid embedding: z_j = (1/|C_j|) Σ_{h∈C_j} h ... a simple and efficient aggregation function agg is performed, which computes the weighted average of all place embeddings, weighted by the number of POIs in each place."

    Substituting Eq. (6) into the weighted average gives L_r = (Σ_j |C_j|·z_j)/(Σ_j |C_j|) = (Σ_j Σ_{h∈C_j} h)/n_r = (1/n_r)Σ_{h∈V_r} h. The region embedding is therefore exactly the ordinary mean of the propagated POI embeddings; the k-means partition, the reduction ratio r, and the cluster count k_r cancel out and have no influence on the 30-dimensional vectors evaluated in Tables 3, 4, and 6. The claimed 'place representation learning' step is thus an identity operation on the final embeddings, and any reported predictive result is the result of mean-pooling SD-CEM+SGC features, not of clustering. The paper's own definitions make the claimed derivation equivalent to its input by construction.

  2. self citation load bearing [Section 4.3.3, after Eq. (6)]
    "It has been theoretically proven [8] that such condensed representations are as expressive as the original POI embeddings, allowing downstream models trained on Z_r to achieve comparable performance to those trained on H_r."

    Reference [8] is the authors' own GECC preprint. It is the sole authority offered for the expressivity guarantee that motivates the clustering step. But because the region aggregator collapses Z_r to a single mean vector before any downstream model is trained, no model is ever trained on Z_r; the cited guarantee about training on condensed graphs cannot justify the reported embeddings. The citation is therefore load-bearing for the paper's methodological premise while being both self-referential and inapplicable to the actual pipeline.

full rationale

The reported numerical comparisons are not fabricated and the underlying SD-CEM + SGC + mean-pooling pipeline is externally evaluated, so there is no fitted-input-called-prediction circularity. However, the paper's central claimed novelty — that clustering POIs into places produces the region embeddings — reduces to a mathematical identity in the paper's own equations. Since the experiments fix the embedding dimension to 30 and describe 'a single meaningful representation for the entire region,' the weighted average of cluster centroids weighted by cluster size is the only reading consistent with Tables 3, 4, and 6, and that average telescopes to the plain mean of the POI embeddings. The clustering step, the reduction ratio r, and all place-level structure have zero effect on the vectors actually evaluated. The GECC self-citation [8] does not rescue the method because the cited expressivity guarantee concerns models trained on condensed graphs, whereas PlaceRep trains no model on the cluster set. Overall this is a partial circularity: the empirical result stands, but the claimed 'place representation learning' derivation is equivalent to mean pooling by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central method rests on pretrained SD-CEM embeddings, a spatial graph construction, SGC propagation, and k-means clustering. The only numbers fitted to the downstream tasks are the per-state, per-task hyperparameters (alpha and r). The paper borrows a theoretical guarantee from the authors' own GECC paper without verifying its conditions, and it introduces the 'place' entity with no independent validation.

free parameters (4)
  • reduction ratio r = 0.02 to 0.1 per state (Table 5)
    Controls number of places per region via k_r = floor(n_r * r); tuned separately for each state and task.
  • alpha_0, alpha_1, alpha_2 = e.g., WY 0.5/0.5/0.0 for population, 0.5/0.0/0.0 for housing
    Propagation weights in the SGC weighted sum (Eq. 3); searched in [0,1] with 0.25 increments per state and task.
  • embedding dimension d = 30
    Set to 30 for all models for fair comparison; affects all downstream results.
  • propagation steps K = 2
    Single-layer SGC with two propagation steps; chosen in implementation details.
assumptions (5)
  • domain assumption SD-CEM category embeddings capture sufficient semantic information for urban prediction
    Section 4.3.1 uses SD-CEM pretrained vectors as the only POI feature; no validation that these embeddings alone are adequate.
  • domain assumption Delaunay triangulation / KNN graphs encode meaningful spatial relationships between POIs
    Section 4.2 adopts these strategies from prior work [5,13,14] without comparing alternatives.
  • ad hoc to paper GECC graph condensation theorem transfers to k-means centroids of SGC-propagated POI embeddings
    Section 4.3.3 cites [8] to claim condensed representations are as expressive, but the theorem's assumptions are not checked for this clustering setting.
  • domain assumption Bisecting k-means clusters correspond to semantically meaningful places
    Section 4.3.3 treats each cluster as a place; evaluation is only qualitative (Figure 5).
  • domain assumption ZIP-code level aggregation with POI-count-weighted averaging is a valid region representation
    Section 4.3.4; no comparison to other aggregators.
invented entities (1)
  • place (cluster of POIs)
    purpose: Intermediate representation between POI and administrative region; used for multi-granular embeddings.
    A place is defined operationally as a k-means cluster; there is no external benchmark validating that these clusters match human-perceived neighborhoods. The qualitative Voronoi inspection (Figure 5) is the only evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PlaceRep: Geospatial Place Representation Learning from Large-Scale Point-of-Interest Data." pith.science (2026). https://pith.science/paper/2AWK7AOJ

@misc{pith2026250702921,
  author       = {Pith},
  title        = {Pith review of: PlaceRep: Geospatial Place Representation Learning from Large-Scale Point-of-Interest Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AWK7AOJ}},
  note         = {Machine review of arXiv:2507.02921}
}
read the original abstract

Learning effective representations of urban environments requires capturing spatial structure beyond fixed administrative boundaries. Existing geospatial representation learning approaches typically aggregate Points of Interest (POIs) into pre-defined administrative regions such as census units or ZIP code areas, assigning a single embedding to each region. However, POIs often form semantically meaningful groups that extend across, within, or beyond these boundaries, defining places that better reflect human activity and urban function. To address this limitation, we propose PlaceRep, a geospatial representation learning method that constructs place-level representations by clustering spatially and semantically related POIs. PlaceRep summarizes large-scale POI graphs from U.S. Foursquare data to produce general-purpose urban region embeddings while automatically identifying places across multiple spatial scales. By eliminating model pre-training, PlaceRep provides a scalable and efficient solution for multi-granular geospatial analysis. Experiments using the tasks of population density estimation and housing price prediction as downstream tasks show that PlaceRep outperforms most state-of-the-art graph-based geospatial representation learning methods and achieves up to a x100 speedup in generating region-level representations on large-scale POI graphs. The implementation of PlaceRep is available at https://github.com/mohammadhashemii/PlaceRep.

Figures

Figures reproduced from arXiv: 2507.02921 by the authors.

Figure 1
Figure 1. The spatial distribution of POIs based on a uniform [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. PlaceRep’s architecture overview. First, it builds POI-level graphs for each state, followed by category feature encoding [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Spatial distribution of absolute housing price es [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Efficiency comparison of region embedding genera [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Voronoi spatial distribution of identified places in ZIP code 30329, Atlanta, GA. Places with the same color belong to [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages

  1. [1]

    Mohit Agarwal, Mimi Sun, Chaitanya Kamath, Arbaaz Muslim, Prithul Sarker, Joydeep Paul, Hector Yee, Marcin Sieniek, Kim Jablonski, Yael Mayer, et al. 2024. General Geospatial Inference with a Population Dynamics Foundation Model. arXiv preprint arXiv:2411.07207(2024)

    arXiv 2024

  2. [2]

    1996.Voronoi diagrams

    Franz Aurenhammer and Rolf Klein. 1996.Voronoi diagrams. Fernuniv., Fach- bereich Informatik

    work page 1996

  3. [3]

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. On the Opportunities and Risks of Foundation Models. arXiv preprint arXiv:2108.07258(2021)

    arXiv 2021

  4. [4]

    Zhifeng Cheng, Jianghao Wang, and Yong Ge. 2022. Mapping monthly population distribution and variation at 1-km resolution across China.International Journal of Geographical Information Science36, 6 (2022), 1166–1184

    work page 2022

  5. [5]

    Boris Delaunay, S Vide, A Lamémoire, and V De Georges. 1934. Bulletin de l’Academie des Sciences de l’URSS.Classe des sciences mathématiques et naturelles 6 (1934), 793–800

    work page 1934

  6. [6]

    Peijun Du, Xuyu Bai, Kun Tan, Zhaohui Xue, Alim Samat, Junshi Xia, Erzhu Li, Hongjun Su, and Wei Liu. 2020. Advances of four machine learning methods for spatial data handling: A review.Journal of Geovisualization and Spatial Analysis 4 (2020), 1–25

    work page 2020

  7. [7]

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. 1996. A density- based algorithm for discovering clusters in large spatial databases with noise. In kdd, Vol. 96. 226–231

    1996

  8. [8]

    Shengbo Gong, Mohammad Hashemi, Juntong Ni, Carl Yang, and Wei Jin

    work page

Show all 42 references
  1. [9]

    Charles R Harris, K Jarrod Millman, Stéfan J Van Der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J Smith, et al. 2020. Array programming with NumPy.nature585, 7825 (2020), 357–362

    2020

  2. [10]

    Mohammad Hashemi, Shengbo Gong, Juntong Ni, Wenqi Fan, B Aditya Prakash, and Wei Jin. 2024. A comprehensive survey on graph reduction: Sparsification, coarsening, and condensation.IJCAI(2024)

    2024

  3. [11]

    Mohammad Hashemi and Andreas Zufle. 2025. From Points to Places: Towards Human Mobility-Driven Spatiotemporal Foundation Models via Understanding Places.Proceedings of the 33rd ACM SIGSPATIAL international conference on advances in geographic information systems(2025)

    2025

  4. [12]

    Mohammad Hashemi and Andreas Züfle. 2026. A Comprehensive Survey of Agentic AI for Spatio-Temporal Data. (2026)

    2026

  5. [13]

    Weiming Huang, Lizhen Cui, Meng Chen, Daokun Zhang, and Yao Yao. 2022. Estimating urban functional distributions with semantics preserved POI embed- ding.International Journal of Geographical Information Science36, 10 (2022), 1905–1930

    2022

  6. [14]

    Weiming Huang, Daokun Zhang, Gengchen Mai, Xu Guo, and Lizhen Cui. 2023. Learning urban region representations with POIs and hierarchical graph infomax. ISPRS Journal of Photogrammetry and Remote Sensing196 (2023), 134–145

    2023

  7. [15]

    Md Ashraful Islam, Mir Mahathir Mohammad, Sarkar Snigdha Sarathi Das, and Mohammed Eunus Ali. 2022. A survey on deep learning based Point-of-Interest (POI) recommendations.Neurocomputing472 (2022), 306–325

    2022

  8. [16]

    Hongwei Jia, Meng Chen, Weiming Huang, Kai Zhao, and Yongshun Gong. 2024. Learning hierarchy-enhanced POI category representations using disentangled mobility sequences. InProceedings of the Thirty-Third International Joint Confer- ence on Artificial Intelligence, Jeju, Korea. 3–9

    2024

  9. [17]

    Ming Jin, Qingsong Wen, Yuxuan Liang, Chaoli Zhang, Siqiao Xue, Xue Wang, James Zhang, Yi Wang, Haifeng Chen, Xiaoli Li, et al. 2023. Large models for time series and spatio-temporal data: A survey and outlook.arXiv preprint arXiv:2310.10196(2023)

    2023 arXiv

  10. [18]

    Zekun Li, Jina Kim, Yao-Yi Chiang, and Muhao Chen. 2022. Spabert: a pre- trained language model from geographic data for geo-entity representation. arXiv preprint arXiv:2210.12213(2022)

    2022 arXiv

  11. [19]

    Yuxuan Liang, Haomin Wen, Yutong Xia, Ming Jin, Bin Yang, Flora Salim, Qing- song Wen, Shirui Pan, and Gao Cong. 2025. Foundation Models for Spatio- Temporal Data Science: A Tutorial and Survey.arXiv preprint arXiv:2503.13502 (2025)

    2025 arXiv

  12. [20]

    Haifeng Niu and Elisabete A Silva. 2021. Delineating urban functional use from points of interest data with neural network embedding: A case study in Greater London.Computers, Environment and Urban Systems88 (2021), 101651

    2021

  13. [21]

    Salma Ommi and Mohammad Hashemi. 2024. Machine learning technique in the north zagros earthquake prediction.Applied Computing and Geosciences22 (2024), 100163

    2024

  14. [22]

    Vinita Rohilla, Sudeshna Chakraborty, Ms Sanika Singh, et al. 2019. Data clus- tering using bisecting k-means. In2019 international conference on computing, communication, and intelligent systems (ICCCIS). IEEE, 80–83

    2019

  15. [23]

    Esther Rolf, Jonathan Proctor, Tamma Carleton, Ian Bolliger, Vaishaal Shankar, Miyabi Ishihara, Benjamin Recht, and Solomon Hsiang. 2021. A generalizable and accessible approach to machine learning with global satellite imagery.Nature communications12, 1 (2021), 4392

    2021

  16. [24]

    Shuoshuo Shang, Shihong Du, Shouji Du, and Shoujie Zhu. 2021. Estimating building-scale population using multi-source spatial data.Cities111 (2021), 103002

    2021

  17. [25]

    Nicolas Tempelmeier, Simon Gottschalk, and Elena Demidova. 2021. GeoVec- tors: a linked open corpus of OpenStreetMap Embeddings on world scale. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 4604–4612

    2021

  18. [26]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. InInternational conference on machine learning. Pmlr, 6861–6871

    2019

  19. [27]

    Lixia Wu, Jianlin Liu, Junhong Lou, Haoyuan Hu, Jianbin Zheng, Haomin Wen, Chao Song, and Shu He. 2023. G2ptl: A pre-trained model for delivery address and its applications in logistics system.arXiv preprint arXiv:2304.01559(2023)

    2023 arXiv

  20. [28]

    Xinhua Wu, Haoyu He, Yanchao Wang, and Qi Wang. 2024. Pretrained mo- bility transformer: A foundation model for human mobility.arXiv preprint arXiv:2406.02578(2024)

    2024 arXiv

  21. [29]

    Yongyang Xu, Bo Zhou, Shuai Jin, Xuejing Xie, Zhanlong Chen, Sheng Hu, and Nan He. 2022. A framework for urban land use classification by integrating the spatial context of points of interest and graph convolutional neural network method.Computers, Environment and Urban System...

    2022

  22. [30]

    Bo Yan, Krzysztof Janowicz, Gengchen Mai, and Song Gao. 2017. From itdl to place2vec: Reasoning about place type similarity and relatedness by learning embeddings from augmented spatial contexts. InProceedings of the 25th ACM SIGSPATIAL international conference on advances in ...

    2017

  23. [31]

    Xiongfeng Yan, Tinghua Ai, Min Yang, and Hongmei Yin. 2019. A graph convolu- tional neural network for classification of building patterns using spatial vector data.ISPRS journal of photogrammetry and remote sensing150 (2019), 259–273

    2019

  24. [32]

    Yao Yao, Xia Li, Xiaoping Liu, Penghua Liu, Zhaotang Liang, Jinbao Zhang, and Ke Mai. 2017. Sensing spatial distribution of urban land use by integrating points- of-interest and Google Word2Vec model.International Journal of Geographical Information Science31, 4 (2017), 825–848

    2017

  25. [33]

    Yao Yao, Xiaoping Liu, Xia Li, Jinbao Zhang, Zhaotang Liang, Ke Mai, and Yatao Zhang. 2017. Mapping fine-scale population distributions at the building level by integrating multisource geospatial big data.International Journal of Geographical Information Science31, 6 (2017), 1220–1244

    2017

  26. [34]

    Jing Yuan, Yu Zheng, and Xing Xie. 2012. Discovering regions of different functions in a city using human mobility and POIs. InProceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining. 186–194

    2012

  27. [35]

    Daochen Zha, Zaid Pervaiz Bhat, Kwei-Herng Lai, Fan Yang, Zhimeng Jiang, Shaochen Zhong, and Xia Hu. 2025. Data-centric artificial intelligence: A survey. Comput. Surveys57, 5 (2025), 1–42

    2025

  28. [36]

    Wei Zhai, Xueyin Bai, Yu Shi, Yu Han, Zhong-Ren Peng, and Chaolin Gu. 2019. Beyond Word2vec: An approach for urban functional region extraction and identification by combining Place2vec and POIs.Computers, environment and urban systems74 (2019), 1–12

    2019

  29. [37]

    Weijia Zhang, Jindong Han, Zhao Xu, Hang Ni, Hao Liu, and Hui Xiong. 2024. Urban foundation models: A survey. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6633–6643

    2024

  30. [38]

    Xiuyuan Zhang, Shihong Du, and Qiao Wang. 2017. Hierarchical semantic cognition for urban functional zones with VHR satellite images and POI data. ISPRS Journal of Photogrammetry and Remote Sensing132 (2017), 170–184

    2017

  31. [39]

    Xiuyuan Zhang, Shihong Du, and Qiao Wang. 2018. Integrating bottom-up clas- sification and top-down feedback for improving urban land-cover and functional- zone mapping.Remote Sensing of Environment212 (2018), 231–248

    2018

  32. [40]

    Shenglin Zhao, Irwin King, and Michael R Lyu. 2016. A survey of point-of- interest recommendation in location-based social networks.arXiv preprint arXiv:1607.00647(2016)

    2016 arXiv

  33. [41]

    Jiong Zhu, Ryan A Rossi, Anup Rao, Tung Mai, Nedim Lipka, Nesreen K Ahmed, and Danai Koutra. 2021. Graph neural networks with heterophily. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 11168–11176

    2021

  34. [2025]

    Scalable Graph Condensation with Evolving Capabilities.arXiv preprint arXiv:2502.17614(2025)

    2025

Pith tools

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