Pith. sign in

REVIEW 4 major objections 4 minor 56 references

POIFormer: A Transformer-Based Framework for Accurate and Scalable Point-of-Interest Attribution

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

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

Pith's one-line read POIFormer claims that POI attribution under realistic GPS noise is best solved by jointly scoring nearby candidates with a causal transformer over a user's past and future visits plus category-level kernel density estimates of crowd visit…

desk verdict A genuinely novel transformer-based POI attribution framework, but the unstated candidate-set construction and missing reproducibility details make the headline claims conditional. read the letter →

arxiv 2507.09137 v1 pith:YC7ZN4WY submitted 2025-07-12 cs.LG cs.AI

Nripsuta Ani Saxena , Shang-Ling Hsu , Mehul Shetty , Omar Alkhadra , Cyrus Shahabi , Abigail L. Horn This is my paper · ORCID
classification cs.LGcs.AI
keywords point-of-interestattributionhumanmobilitytransformerGPSnoisekerneldensityestimationstaypointtrajectorymodeling
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

POI attribution—deciding which store, cafe, or park a noisy GPS stay actually corresponds to—is a foundational step for mobility analytics, and it fails when POIs are packed more tightly than the GPS error (2–20 m). This paper claims that the right model for the task is one that scores each nearby candidate POI using both an individual's trajectory context and crowd-level visit patterns, rather than proximity, polygons, or single-category assumptions alone. The proposed POIFormer pairs a causal transformer that reads the user's past and future visits (with the current stay masked) with precomputed category-level kernel density estimates of when and where people visit. On the dense, noisy NUMOSIM simulation of Los Angeles, it reports top-1 accuracy of 0.7223 versus 0.7023 for the SafeGraph pipeline, and much larger top-3 and top-5 margins; it wins five of the six noisy experiments overall. If true, this makes accurate attribution attainable without proprietary building-footprint data layers.

What carries the argument

The load-bearing object is the factorized scoring identity in Equation 6, which turns attribution into a sum over the candidate POI's categories of a learned log-prior from the transformer context and a precomputed KDE log-likelihood. The sequence encoder uses Space2Vec for coordinates and Time2Vec for arrival/departure times, feeds the concatenated stay embeddings through a causal transformer encoder adapted from TrajGPT, and masks the location of the current stay so the model must infer it from past and future visits. The spatiotemporal likelihood module queries category-specific KDEs trained offline on historical stay data. This factorization is what lets the model scale to large POI sets and avoid the single-category assumption of prior probabilistic models.

What would settle it

Construct the candidate set $P'$ using the paper's actual shortlisting rule and radius (which the paper does not state), then measure recall = fraction of stays whose true POI is in $P'$; if that recall is below the reported top-1 accuracy, the reported accuracy cannot be achieved by the stated pipeline. A concrete check is to rerun the NUMOSIM comparison with $P'$ built at radii such as 25 m, 50 m, and 100 m and show that top-1 accuracy tracks $P'$ recall.

Watch

Extended reading notes

Core claim

POIFormer's central claim is that the probability a stay belongs to a candidate POI factorizes over the POI's semantic categories: $\log \Pr(p \mid t, l, H) \approx \sum_{c \in C(p)} [\log \Pr(t,l \mid c) + \log \Pr(c \mid H)]$. Here $\Pr(c \mid H)$ is a learned category prior produced by a causal transformer over the user's trajectory—both past and future visits, with the current visit's location masked—and $\Pr(t,l \mid c)$ is a category-level kernel density estimate of the joint distribution of location and time of visits, precomputed from historical stay data. The model is trained end-to-end to minimize the negative log likelihood of the true POI. On the NUMOSIM dense-urban dataset with added GPS noise, POIFormer reports top-1/top-3/top-5 accuracy of 0.7223/0.9093/0.9653 against SafeGraph's 0.7023/0.7938/0.8388; on the lower-density Breadcrumbs dataset with noise it trails SafeGraph by 1.15 points in top-1 but leads by 2.96 and 2.68 points in top-3 and top-5. The paper interprets these results as showing that jointly modeling individual preferences, crowd behavior, POI semantics, and spatiotemporal context makes attribution stable under GPS noise and dense clustering without relying on hard-to-access spatial data layers.

Load-bearing premise

The method only ever picks from a pre-chosen shortlist of nearby POIs; the paper does not state the radius or rule used to build that shortlist, and if the real destination is left off it, accuracy is zero regardless of the model.

Editorial extensions

If this is right

  • Attribution accuracy can be maintained under realistic smartphone GPS noise without building-footprint polygons or spatial-hierarchy metadata, so the method transfers to regions where such data layers do not exist.
  • In dense urban settings the biggest gains come in the top-3/top-5 rankings: about 14.6 and 15.1 percentage points over the SafeGraph pipeline on NUMOSIM, which matters for downstream tasks that consume ranked candidate sets.
  • Both modeling components carry weight: ablations on 100,000 NUMOSIM stays show removing the KDEs costs 2.31 points of top-1 accuracy, and removing the learned category prior costs 4.04 points top-1 and 20.72 points top-3.
  • The same model that leads in dense LA keeps near-identical top-1 accuracy on Breadcrumbs with and without added noise (0.9188 vs 0.9169), suggesting the learned trajectory context is stable even where the KDE and category prior have little signal.

Reading between the lines

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

  • Beyond the paper, candidate-set construction is the hidden bottleneck: a deployment must measure the recall of its shortlisting rule, and accuracy should be reported conditioned on whether the true POI made the shortlist.
  • Beyond the paper, because the scoring sum runs over each POI's category list, the method's ceiling is tied to category-taxonomy quality; coarse or missing categories would silently cap achievable accuracy.
  • Beyond the paper, the dense-urban evaluation is synthetic; real dense-city traces with ground-truth labels would verify whether the margin over polygon-based attribution survives outside simulation.
  • Beyond the paper, the KDE crowd priors are static; time-varying or seasonal prior updates are a testable extension for long-running deployments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces POIFormer, a transformer-based framework for point-of-interest (POI) attribution from GPS stay points. The model combines an individual-level transformer encoder that uses past and future visits, with category-level kernel density estimates that capture crowd-level spatiotemporal visit patterns. The final score for a candidate POI is a sum over the POI's categories of a learned category prior and a KDE log-likelihood. The authors evaluate on two public datasets, NUMOSIM and Breadcrumbs, with and without added Gaussian noise, and report top-1, top-3, and top-5 accuracy against closest-centroid, Nishida et al., and SafeGraph baselines. They claim that POIFormer outperforms all baselines in five of six noisy-data experiments, with particularly large top-3 and top-5 margins on NUMOSIM, and they provide an ablation study on the NUMOSIM subset.

Significance. If the reported results hold, POIFormer would be a practical contribution to POI attribution: it avoids reliance on proprietary polygon and hierarchy data, handles multi-category POIs, and maintains accuracy under GPS noise in dense urban settings. The use of publicly available datasets is a strength, and the factorization into a learned category prior and precomputed KDE likelihoods is transparent and efficient. However, the central empirical claim is currently not fully verifiable because the candidate-POI set is not operationalized, statistical uncertainty is not reported, and the SafeGraph baseline implementation is underspecified. These issues are fixable, but they are load-bearing for the claimed superiority.

major comments (4)
  1. [§3.2 and §4.1.2] The candidate set P' is defined only as 'feasible POIs near the stay location' (Section 3.2), and the experimental section never states how P' is constructed for either dataset. Since Eq. (6) and the top-k metrics are evaluated only over P', every reported accuracy is conditioned on the true POI being included in P'; if the true POI is filtered out, the model scores zero by construction. The absence of a radius, a scoring rule, or a recall check makes the headline comparisons in Tables 2 and 3 unverifiable and potentially unfair relative to SafeGraph, which uses polygon-based joins. The authors should specify the candidate-generation procedure (including distance thresholds, category filters, or maximum candidate counts) for NUMOSIM and Breadcrumbs, report the recall of P', and confirm that the same candidate sets are used for all baselines.
  2. [§4.2.1, Tables 2-3] The central empirical claim that POIFormer outperforms baselines in five of six noisy settings rests on single-run accuracy values with no confidence intervals, no repeated seeds, and no significance tests. The decisive comparisons are often small (e.g., NUMOSIM Top-1, 0.7223 vs 0.7023; Breadcrumbs noisy Top-1, 0.9169 vs 0.9276), and the ablation in Table 4 uses a subset of 100,000 stays without any variance estimate. Reporting mean and standard deviation over multiple runs, or at least confidence intervals, is necessary before the claimed 'consistent' advantage can be assessed.
  3. [§3.2.1, Eq. (4)] The product over categories in Eq. (4) may systematically penalize multi-category POIs: because each factor is a probability (or, in log space, a negative contribution), a POI with more categories receives a lower score than a single-category POI with the same per-category agreement, unless the KDE density values are large enough to compensate. This is in tension with the paper's stated motivation that POIFormer handles multi-functional POIs, especially since NUMOSIM POIs are described as mostly having three or more categories. The authors should justify this aggregation, normalize it, or provide evidence that the bias does not drive the comparison.
  4. [§4.1.4] The SafeGraph baseline, which is the strongest competitor, is described only by a reference to a white paper; no implementation details are given for how its polygon and hierarchy data layers were obtained or used in the reported runs. Since the paper's contribution is partly framed as avoiding such proprietary layers, the comparison should state whether the SafeGraph numbers come from the vendor pipeline or from a reimplementation, and if the latter, which components were simplified. This is needed for the reader to judge the fairness of the reported 14.55% and 15.08% Top-3/Top-5 margins on NUMOSIM.
minor comments (4)
  1. [§4.2.1] The reported '14.55%' and '15.08%' gains are relative improvements over the SafeGraph baseline, not absolute percentage-point differences; for example, 0.9093 vs 0.7938 is 11.55 percentage points in Top-3 accuracy. Please state both the relative and absolute differences to avoid ambiguity.
  2. [§4.3] The ablation text mixes percentage-point and relative-change language: 'drops by 2.31%' from 63.30% to 60.99% is 2.31 percentage points, and the same convention should be used for the Top-3 and Top-5 decreases.
  3. [§4.2.1] The sentence 'POIFormer outperforms all baselines in five out of six noisy-data experiments' is correct only if Top-1, Top-3, and Top-5 on each dataset are counted as separate experiments; the Breadcrumbs noisy Top-1 loss to SafeGraph should be acknowledged in the same sentence for precision.
  4. [§3.2.6] The paper does not report the KDE bandwidth selection procedure or the resulting bandwidth values for each category; adding these details, along with the distribution of candidate set sizes, would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: POIFormer is a supervised classifier whose components are fit on training labels and evaluated on held-out stays.

full rationale

The derivation chain in Sections 3.2.1-3.2.8 is a standard supervised factorization: Eq. (6) defines each candidate score as a sum over categories of log Pr(t,l|c) plus log Pr(c|H). The category prior is produced by a transformer trained end-to-end with cross-entropy against ground-truth labels (Eqs. 7-8), and the KDE likelihoods are fit on historical stay data. Neither component is defined in terms of the test labels or the reported accuracy, and the reported top-k numbers are computed on held-out staypoints from NUMOSIM and Breadcrumbs. The self-citations (TrajGPT for the encoder architecture, NUMOSIM as a dataset) are building blocks and are not used to define the target result. The main validity concern is that the candidate set P' (Section 3.2) is only described as 'feasible POIs near the stay location' with no radius or recall check, so every accuracy number is bounded by the recall of P'; this is an experimental-validity gap, not a circularity, because the paper gives no indication that P' is constructed from the ground-truth POI or from the fitted model's own outputs. No equation or reported result reduces to its own inputs by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard supervised-learning machinery plus five domain assumptions. The most consequential unreported free parameters are the candidate set radius and KDE bandwidth, both of which directly bound the reported accuracies. No entity is invented beyond the learned weights.

free parameters (4)
  • KDE bandwidth per POI category = not reported
    The category-specific KDEs in Section 3.2.2 are the crowd likelihood engine, but the bandwidth selection rule is never stated; top-k results depend on it.
  • Candidate POI set radius = not reported
    Section 3.2 defines feasible candidates as POIs 'near' the stay but never gives a radius or score threshold, which bounds achievable accuracy.
  • Coordinate noise augmentation scale = sigma in {0.0002, 0.0001, 0.00005}
    Section 4.1.2 adds Gaussian noise to stay coordinates; the paper says differences across sigma are not significant but provides no detailed breakdown.
  • Transformer and optimizer hyperparameters = not reported
    Hidden size, layers, heads, learning rate, batch size, and sequence length are omitted, so the exact model is not reproducible.
assumptions (5)
  • domain assumption Each POI has a known, fixed set of semantic categories C(p).
    Used in Definitions 3.1 and Eq. 4; in real databases category labels are often missing, noisy, or multi-label and inconsistent across sources.
  • ad hoc to paper Pr(t,l|c,H) is approximated by Pr(t,l|c), i.e., the spatiotemporal distribution given a category does not depend on the user's history.
    Equation 3 in Section 3.2.1; this conditional independence collapses two sources of evidence into one and may fail for users with strong personal routines.
  • domain assumption A POI's categories contribute independent evidence, so likelihoods multiply across categories.
    Equation 4; treats categories of a multi-category POI as conditionally independent, which can overcount overlapping semantics like 'cafe' and 'restaurant'.
  • domain assumption Category-level KDEs trained on historical stays transfer to the test period and to other regions.
    Section 3.2.2 and Section 4.1; NUMOSIM uses separate four-week train/test periods and Breadcrumbs uses a different city, so stationarity of visit patterns is assumed.
  • domain assumption At inference, both past and future visits of the target stay are available.
    Section 3.2.3 masks only the target location; this is fine for offline attribution of historical logs but not for real-time attribution during a trip.

how reviews work

0 comments
Cite this review

Pith. "Pith review of POIFormer: A Transformer-Based Framework for Accurate and Scalable Point-of-Interest Attribution." pith.science (2026). https://pith.science/paper/YC7ZN4WY

@misc{pith2026250709137,
  author       = {Pith},
  title        = {Pith review of: POIFormer: A Transformer-Based Framework for Accurate and Scalable Point-of-Interest Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YC7ZN4WY}},
  note         = {Machine review of arXiv:2507.09137}
}
read the original abstract

Accurately attributing user visits to specific Points of Interest (POIs) is a foundational task for mobility analytics, personalized services, marketing and urban planning. However, POI attribution remains challenging due to GPS inaccuracies, typically ranging from 2 to 20 meters in real-world settings, and the high spatial density of POIs in urban environments, where multiple venues can coexist within a small radius (e.g., over 50 POIs within a 100-meter radius in dense city centers). Relying on proximity is therefore often insufficient for determining which POI was actually visited. We introduce \textsf{POIFormer}, a novel Transformer-based framework for accurate and efficient POI attribution. Unlike prior approaches that rely on limited spatiotemporal, contextual, or behavioral features, \textsf{POIFormer} jointly models a rich set of signals, including spatial proximity, visit timing and duration, contextual features from POI semantics, and behavioral features from user mobility and aggregated crowd behavior patterns--using the Transformer's self-attention mechanism to jointly model complex interactions across these dimensions. By leveraging the Transformer to model a user's past and future visits (with the current visit masked) and incorporating crowd-level behavioral patterns through pre-computed KDEs, \textsf{POIFormer} enables accurate, efficient attribution in large, noisy mobility datasets. Its architecture supports generalization across diverse data sources and geographic contexts while avoiding reliance on hard-to-access or unavailable data layers, making it practical for real-world deployment. Extensive experiments on real-world mobility datasets demonstrate significant improvements over existing baselines, particularly in challenging real-world settings characterized by spatial noise and dense POI clustering.

Figures

Figures reproduced from arXiv: 2507.09137 by the authors.

Figure 1
Figure 1. Illustration of key concepts in mobility analysis: Figure 1a showcases a GPS trace recorded from an individual’s [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The architecture of POIFormer. characteristics align with the aggregated behavioral patterns ob￾served for category 𝑐, allowing the model to downweight unlikely POI candidates accordingly. 3.2.7 Scoring Candidate POIs . To compute the final score for each candidate POI 𝑝 ∈ 𝑃 ′ , we combine the context-based prior and the spatiotemporal likelihood according to the formulation in Section 3.2.1. Since the overall proba… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 53 canonical work pages

  1. [1]

    Gabriel Agamennoni, Juan Nieto, and Eduardo Nebot. 2009. Mining GPS data for extracting significant places. In 2009 IEEE International Conference on Robotics and Automation. IEEE, 855–862

    work page 2009

  2. [2]

    Daniel Ashbrook and Thad Starner. 2003. Using GPS to learn significant locations and predict movement across multiple users. Personal and Ubiquitous computing 7 (2003), 275–286

    work page 2003

  3. [3]

    Christine Bauer and Christine Strauss. 2016. Location-based advertising on mobile devices: A literature review and analysis. Management review quarterly 66, 3 (2016), 159–194

    work page 2016

  4. [4]

    Xin Cao, Gao Cong, and Christian S Jensen. 2010. Mining significant semantic locations from GPS data. Proceedings of the VLDB Endowment 3, 1-2 (2010), 1009–1020

    work page 2010

  5. [5]

    Joon Wayn Cheong, Binghao Li, Andrew G Dempster, and Chris Rizos. 2009. GPS/WiFi real-time positioning device: An initial outcome. In Location Based Services and TeleCartography II: From sensor fusion to context models . Springer, 439–456

    work page 2009

  6. [6]

    Gérard Cliquet and Jérôme Baray. 2020. Location-based marketing: geomarketing and geolocation. John Wiley & Sons

    work page 2020

  7. [7]

    Andrea Cuttone, Sune Lehmann, and Jakob Eg Larsen. 2014. Inferring human mobility from sparse low accuracy mobile sensing data. In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication. 995–1004

    work page 2014

  8. [8]

    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

Show all 56 references
  1. [9]

    Federal Highway Administration. 2022. 2022 NextGen National Household Travel Survey Core Data. https://nhts.ornl.gov. Available online: https://nhts.ornl.gov

    2022

  2. [10]

    Pedro J Fernández, José Santa, and Antonio F Skarmeta. 2020. Hybrid positioning for smart spaces: proposal and evaluation. Applied Sciences 10, 12 (2020), 4083

    2020

  3. [11]

    Simon Gröchenig and Cornelia Schneider. 2016. A Cookie-Cutter Approach for Determining Places and Stays from Movement Data. GI_Forum 1 (2016), 53–64

    2016

  4. [12]

    Ramaswamy Hariharan and Kentaro Toyama. 2004. Project Lachesis: parsing and modeling location histories. InInternational Conference on Geographic Information Science. Springer, 106–124

    2004

  5. [13]

    Shang-Ling Hsu, Emmanuel Tung, John Krumm, Cyrus Shahabi, and Khurram Shafique. 2024. Trajgpt: Controlled synthetic trajectory generation using a multitask transformer-based spatiotemporal model. In Proceedings of the 32nd ACM International Conference on Advances in Geographic...

    2024

  6. [14]

    Seyed Mehran Kazemi, Rishab Goel, Sepehr Eghbali, Janahan Ramanan, Jaspreet Sahota, Sanjay Thakur, Stella Wu, Cathal Smyth, Pascal Poupart, and Marcus Brubaker. 2019. Time2vec: Learning a vector representation of time. arXiv preprint arXiv:1907.05321 (2019)

    2019 arXiv

  7. [15]

    Minkyong Kim, David Kotz, and Songkuk Kim. 2006. Extracting a mobility model from real user traces. (2006)

    2006

  8. [16]

    John Krumm and Dany Rouhana. 2013. Placer: semantic place labels from diary data. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing. 163–172

    2013

  9. [17]

    Kari Laasonen, Mika Raento, and Hannu Toivonen. 2004. Adaptive on-device location recognition. InInternational Conference on Pervasive Computing. Springer, 287–304

    2004

  10. [18]

    Qingchun Li, Liam Bessell, Xin Xiao, Chao Fan, Xinyu Gao, and Ali Mostafavi

  11. [19]

    Defu Lian and Xing Xie. 2011. Learning location naming from user check-in histories. In Proceedings of the 19th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems . 112–121

    2011

  12. [20]

    Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, and Ni Lao. 2020. Multi-scale representation learning for spatial feature distributions using grid cells. arXiv preprint arXiv:2003.00824 (2020)

    2020 arXiv

  13. [21]

    Chuishi Meng, Yu Cui, Qing He, Lu Su, and Jing Gao. 2017. Travel purpose inference with GPS trajectories, POIs, and geo-tagged social media data. In 2017 IEEE International Conference on Big Data (Big Data) . IEEE, 1319–1324

    2017

  14. [22]

    Krista Merry and Pete Bettinger. 2019. Smartphone GPS accuracy study in an urban environment. PloS one 14, 7 (2019), e0219890

    2019

  15. [23]

    Marian Mohr, Christopher Edwards, and Ben McCarthy. 2008. A study of LBS accuracy in the UK and a novel approach to inferring the positioning technology employed. Computer Communications 31, 6 (2008), 1148–1159

    2008

  16. [24]

    Esmond Mok and Günther Retscher. 2007. Location determination using WiFi fingerprinting versus WiFi trilateration. Journal of Location Based Services 1, 2 (2007), 145–159

    2007

  17. [25]

    Esmond Mok, Guenther Retscher, and Chen Wen. 2012. Initial test on the use of GPS and sensor data of modern smartphones for vehicle tracking in dense high rise environments. In 2012 Ubiquitous Positioning, Indoor Navigation, and Location Based Service (UPINLBS) . IEEE, 1–7

    2012

  18. [26]

    Arielle Moro, Vaibhav Kulkarni, Pierre-Adrien Ghiringhelli, Bertil Chapuis, Kévin Huguenin, and Benoît Garbinato. 2019. Breadcrumbs: a rich mobility dataset with point-of-interest annotations. In Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Ge...

    2019

  19. [27]

    Asya Natapov, Achituv Cohen, and Sagi Dalyot. 2024. Urban planning and design with points of interest and visual perception. Environment and Planning B: Urban Analytics and City Science 51, 3 (2024), 641–655

    2024

  20. [28]

    National Coordination Office for Space-Based Positioning, Navigation, and Tim- ing. n.d.. GPS Accuracy. https://www.gps.gov/systems/gps/performance/ accuracy/ Accessed: 2025-05-27

    2025

  21. [29]

    Kyosuke Nishida, Hiroyuki Toda, Takeshi Kurashima, and Yoshihiko Suhara

  22. [30]

    Petteri Nurmi and Sourav Bhattacharya. 2008. Identifying meaningful places: The non-parametric way. In International Conference on Pervasive Computing . Springer, 111–127

    2008

  23. [31]

    Andrey Tietbohl Palma, Vania Bogorny, Bart Kuijpers, and Luis Otavio Alvares

  24. [32]

    Rafael Pérez-Torres, César Torres-Huitzil, and Hiram Galeana-Zapién. 2016. Full on-device stay points detection in smartphones for location-based mobile appli- cations. Sensors 16, 10 (2016), 1693

    2016

  25. [33]

    Jan Petzold, Faruk Bagci, Wolfgang Trumler, and Theo Ungerer. 2006. Compari- son of different methods for next location prediction. In Euro-Par 2006 Parallel Processing: 12th International Euro-Par Conference, Dresden, Germany, August 28–September 1, 2006. Proceedings 12 . Spr...

    2006

  26. [34]

    Achilleas Psyllidis, Song Gao, Yingjie Hu, Eun-Kyeong Kim, Grant McKenzie, Ross Purves, May Yuan, and Clio Andris. 2022. Points of Interest (POI): A commentary on the state of the art, challenges, and prospects for the future. Computational urban science 2, 1 (2022), 20

    2022

  27. [35]

    SafeGraph. 2025. Determining Points of Interest Visits From Location Data: A Technical Guide to Visit Attribution. https://www.safegraph.com/guides/visit- attribution-white-paper. Places Data Curated for Accurate Geospatial Analytics

    2025

  28. [36]

    Salvatore Scellato, Mirco Musolesi, Cecilia Mascolo, Vito Latora, and Andrew T Campbell. 2011. Nextplace: a spatio-temporal prediction framework for pervasive systems. In Pervasive Computing: 9th International Conference, Pervasive 2011, San Francisco, USA, June 12-15, 2011. P...

    2011

  29. [37]

    Elan Shore. 2022. 5 Standout Stats About Downtown LA. https: //downtownla.com/article/value-of-downtowns#:~:text=DowntownâĂŹs% 20outsized%20economic%20contribution%20is,the%20rest%20of%20the%20city

    2022

  30. [38]

    Maria Despoina Siampou, Jialiang Li, John Krumm, Cyrus Shahabi, and Hua Lu. 2025. Poly2Vec: Polymorphic Fourier-Based Encoding of Geospatial Objects for GeoAI Applications. In Forty-second International Conference on Machine Learning

    2025

  31. [39]

    Martin L Sollie, Kristoffer Gryte, Torleiv H Bryne, and Tor Arne Johansen. 2022. Outdoor navigation using bluetooth angle-of-arrival measurements. IEEE Access 10 (2022), 88012–88033

    2022

  32. [40]

    Chris Stanford, Suman Adari, Xishun Liao, Yueshuai He, Qinhua Jiang, Chenchen Kuai, Jiaqi Ma, Emmanuel Tung, Yinlong Qian, Lingyi Zhao, et al . 2024. NU- MOSIM: A Synthetic Mobility Dataset with Anomaly Detection Benchmarks. In Proceedings of the 1st ACM SIGSPATIAL Internation...

    2024

  33. [41]

    Jun Suzuki, Yoshihiko Suhara, Hiroyuki Toda, and Kyosuke Nishida. 2019. Per- sonalized visited-poi assignment to individual raw GPS trajectories. ACM Trans- actions on Spatial Algorithms and Systems (TSAS) 5, 3 (2019), 1–28

    2019

  34. [42]

    Nattapong Swangmuang and Prashant Krishnamurthy. 2008. An effective lo- cation fingerprint model for wireless indoor localization. Pervasive and Mobile Computing 4, 6 (2008), 836–850

    2008

  35. [43]

    Verena Venek, Richard Brunauer, and Cornelia Schneider. 2016. Evaluating the Brownian bridge movement model to determine regularities of people’s movements. GI_Forum 2 (2016), 20–35

    2016

  36. [44]

    Michael Wallbaum. 2007. A priori error estimates for wireless local area network positioning systems. Pervasive and Mobile Computing 3, 5 (2007), 560–580

    2007

  37. [45]

    Dongbo Xi, Fuzhen Zhuang, Yanchi Liu, Jingjing Gu, Hui Xiong, and Qing He

  38. [46]

    Hao Xue, Flora Salim, Yongli Ren, and Nuria Oliver. 2021. MobTCast: Leveraging auxiliary trajectory forecasting for human mobility prediction. Advances in Neural Information Processing Systems 34 (2021), 30380–30391

    2021

  39. [47]

    Di Yao, Chao Zhang, Jianhui Huang, and Jingping Bi. 2017. Serm: A recurrent model for next location prediction in semantic trajectories. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management . 2411–2414

    2017

  40. [48]

    Paul A Zandbergen. 2009. Accuracy of iPhone locations: A comparison of assisted GPS, WiFi and cellular positioning. Transactions in GIS 13 (2009), 5–25. , , Nripsuta Ani Saxena, Shang-Ling Hsu, Mehul Shetty, Omar Alkhadra, Cyrus Shahabi, and Abigail L. Horn

    2009

  41. [49]

    Keshu Zhang, Haifeng Li, Kari Torkkola, and Mike Gardner. 2007. Adaptive learning of semantic locations and routes. InInternational Symposium on Location- and Context-A wareness. Springer, 193–210

    2007

  42. [50]

    Zheng Zhang, Hossein Amiri, Dazhou Yu, Yuntong Hu, Liang Zhao, and Andreas Züfle. 2024. Transferable Unsupervised Outlier Detection Framework for Human Semantic Trajectories. In Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems ...

    2024

  43. [51]

    Yu Zheng and Xiaofang Zhou. 2011. Computing with spatial trajectories. Springer Science & Business Media

    2011

  44. [52]

    home", "work

    Changqing Zhou, Dan Frankowski, Pamela Ludford, Shashi Shekhar, and Loren Terveen. 2007. Discovering personally meaningful places: An interactive clus- tering approach. ACM Transactions on Information Systems (TOIS) 25, 3 (2007), 12–es. POIFormer: A Transformer-Based Framework...

    2007

  45. [2008]

    In Proceedings of the 2008 ACM symposium on Applied computing

    A clustering-based approach for discovering interesting places in trajecto- ries. In Proceedings of the 2008 ACM symposium on Applied computing . 863–868

    2008

  46. [2014]

    In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing

    Probabilistic identification of visited point-of-interest for personalized automatic check-in. In Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing . 631–642

    2014

  47. [2019]

    In Proceedings of the AAAI conference on artificial intelligence (2019) , Vol

    Modelling of bi-directional spatio-temporal dependence and users’ dynamic preferences for missing poi check-in identification. In Proceedings of the AAAI conference on artificial intelligence (2019) , Vol. 33. 5458–5465

    2019

  48. [2021]

    Royal Society open science 8, 1 (2021), 201209

    Disparate patterns of movements and visits to points of interest located in urban hotspots across US metropolitan cities during COVID-19. Royal Society open science 8, 1 (2021), 201209

    2021

Pith tools

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