Pith. sign in

REVIEW 4 major objections 5 minor 60 references

CoBAD: Modeling Collective Behaviors for Human Mobility Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Collective mobility anomalies are detectable from missing companions, CoBAD shows.

desk verdict CoBAD is a serious new problem formulation with a sensible model, but the ghost node mechanism is underspecified and the evaluation is synthetic-only; send to review, expect heavy revision. read the letter →

arxiv 2508.14088 v1 pith:PJPLHGF2 submitted 2025-08-13 cs.LG cs.AIcs.SI

Haomin Wen , Shurui Cao , Leman Akoglu This is my paper
classification cs.LGcs.AIcs.SI
keywords collectiveanomalydetectionhumanmobilityspatiotemporalmodelinggraphtransformerco-occurrenceeventself-supervisedpretrainingabsencesequences
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that mobility anomalies arising from relationships between people—not just an individual's unusual movements—can be detected automatically without labels. It argues that prior work stops at individual behavior, so it misses cases like a child alone at home when a parent is usually present. CoBAD models each person's stay-event sequence together with a co-occurrence event graph, and learns normal collective behavior via masked reconstruction of both events and links. On two large simulated mobility datasets, the authors report event-level gains of 13–18% in AUCROC and 19–70% in AUCPR over baselines, and they claim this is the first method to detect absence-type collective anomalies. A sympathetic reader should care because public safety and urban-planning monitoring often depend on exactly these joint behavioral expectations.

What carries the argument

The central machinery is the two-stage attention (TSA) layer: cross-time self-attention runs along each individual's event sequence to capture personal spatiotemporal patterns, and cross-people attention runs over a co-occurrence event graph whose nodes are stay events and whose edges connect events within 40 meters with overlapping time intervals. Masked pre-training reconstructs both masked event features and masked links; at inference, a ghost-node mechanism lets the model score links to people who are absent from an event, enabling absence-anomaly detection.

What would settle it

Take a real mobility dataset with known ground-truth relationships (e.g., household members or declared friends), inject absence anomalies by removing one expected co-occurring person, and measure CoBAD's event-level AUCPR. If it does not beat a simple historical-co-occurrence-frequency baseline on that real data, the co-occurrence signal is not carrying the detection. A second direct check: manually inspect events CoBAD flags as missing-link anomalies and verify whether the supposedly missing person was actually elsewhere; frequent false ghost links would show the 40-meter threshold is too lo

Watch

Extended reading notes

Core claim

On its own terms, CoBAD's central claim is that collective anomalies in human mobility are learnable in an unsupervised way if the model is given, for each target person, the event sequences of related people and a graph of co-occurrence links (events within 40 meters with overlapping time intervals). The model's two-stage attention first encodes temporal dependencies within each sequence, then aggregates information across people through graph attention. Pre-training asks the model to reconstruct masked event attributes and masked links; at test time, an event is scored by the maximum of three signals: how poorly its features reconstruct, how unexpected its observed links are, and how likel

Load-bearing premise

The whole method assumes that two people whose stay events are within 40 meters with overlapping time intervals are behaviorally connected; if proximity does not imply a meaningful relationship in real data, the learned 'collective behavior' is noise and the anomaly scores will flag false positives and miss true absences.

Editorial extensions

If this is right

  • Without any labeled anomalies, CoBAD can flag an event as anomalous because someone who normally should be there is absent, a type prior methods miss.
  • The observed-link score can catch unexpected co-occurrences, such as a stranger joining a family gathering.
  • Anomaly scores decompose into interpretable components, so each flag carries a type: individual deviation, unexpected link, or missing link.
  • On the two simulated datasets, event-level AUCROC improves by 13–18% and AUCPR by 19–70% over the strongest baseline; agent-level gains are smaller but still positive.
  • The learned co-occurrence model outperforms simple historical-frequency and duration heuristics on link prediction.

Reading between the lines

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

  • A testable extension: feed the same two-stage architecture with relationship signals beyond spatial co-occurrence—household membership, call records, or explicit social ties—to see whether absence detection improves on real-world data.
  • The ghost-node mechanism suggests a general template for scoring 'missing interaction' anomalies whenever absence is observable only through expected links; one could transfer it to fraud rings or coordinated teams.
  • Because the evaluation uses simulated data with anomalies injected to match the co-occurrence definition, real deployment would benefit from validation against independently labeled real-world absences, such as known hospitalizations or travel separations.
  • If the 40-meter, time-overlap threshold were replaced by learned or context-dependent interaction rules, the same scoring function might reduce false positives from mere proximity in crowded places.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CoBAD, an unsupervised model for collective anomaly detection in human mobility. It formalizes collective event sequences (CES) and a co-occurrence event graph, then uses a two-stage attention mechanism (cross-time and cross-people) with masked pre-training to reconstruct both event attributes and co-occurrence links. At inference, three anomaly scores are combined: node reconstruction error, unexpected observed links, and missing/absence links; the latter is computed via a proposed 'ghost node' mechanism. Experiments on two simulated Tokyo mobility datasets (MobilitySim-A/B) with injected anomalies report large AUCROC/AUCPR improvements over several baselines, and ablations show the contribution of each score component.

Significance. If the claims hold, the paper addresses a genuine and underexplored problem: event-level collective anomaly detection in human mobility, especially absence-type anomalies. The two-stage attention design is a reasonable way to combine intra-person and inter-person dependencies, and the public code release is a strength. The ablation study and case studies help interpretability. However, the significance is substantially tempered by two issues: the absence-anomaly mechanism—the paper's main novel contribution—is under-specified to the point of being unreproducible, and the evaluation is on simulated data whose anomaly injection closely mirrors the model's co-occurrence definition. The reported improvements are against weak or mismatched baselines, so the headline performance claim is not yet fully supported.

major comments (4)
  1. [Sec. 4.2 / Alg. 2] The 'ghost node' mechanism underlying Eq. (12) is not specified. Algorithm 2 line 1 only says 'Add ghost node for sequences in T_u^w'. The paper does not state how many ghost nodes are added, what features they receive, how candidate missing neighbors are selected (all non-neighbors? a sampled subset?), or how they are inserted into the input tensor X and the cross-people attention of Eq. (4). Because Eq. (12) is a max over s not in N(e), the candidate set must be defined; otherwise the score is not computable. The reported CoBAD-not-obs-link result (AUCROC 0.717, Fig. 6) is therefore not reproducible from the text. Please provide a complete algorithmic specification, including feature construction and graph insertion.
  2. [Sec. 5.1.2 / Table 2] The central performance claim of 13%-18% AUCROC improvement is made against a limited baseline set. Graph-based collective anomaly detection methods from the related work (e.g., dynamic graph link-anomaly methods) are excluded with a one-sentence justification, and the only link-aware baseline, TransformerLink-AD, uses historical meeting frequency rather than a learned model. The trajectory baselines cannot produce event-level scores at all. Thus Table 2 compares CoBAD against the nearest available baselines, not against state-of-the-art collective anomaly detectors. Please either include adapted graph-based collective AD baselines that use spatiotemporal event features, or state clearly that the improvement is relative to these particular baselines.
  3. [Sec. 5.1.1 / Sec. 5.2] The evaluation uses only two simulated datasets, and the injected anomalies are generated precisely by changing co-occurrence relationships: unexpected occurrence modifies location/POI to create co-occurrence, absence removes a participant from a co-occurrence, and synthetic coordination creates artificial co-occurrence. Since the model's graph and link scores are defined by the same co-occurrence relation, the evaluation is aligned with the model in a way that may not reflect real-world collective behaviors. The paper's own limitation statement admits there is no real labeled data. Please add an external-validity analysis, e.g., a sensitivity study with different delta thresholds or a qualitative evaluation on a real unlabeled dataset, and discuss how false positives would arise if co-occurrence does not imply a meaningful social link.
  4. [Eqs. (11)-(12)] The link score S_link used at inference is never formally defined. Eq. (8) defines sim(s,d) inside the contrastive loss, but Eqs. (11)-(12) assume an S_link that is not specified. If S_link is the raw cosine similarity, its scale is not calibrated for taking a max over a large candidate set; if it is a normalized probability, the normalization depends on the negative-set construction, which is also unspecified. Please define S_link precisely and explain how it is derived from the pre-trained model.
minor comments (5)
  1. [Abstract / Table 2] The abstract states improvements of 13%-18% AUCROC and 19%-70% AUCPR without noting that these are event-level improvements. Agent-level gains in Table 2 are much smaller (2.3%-10.3%). Please qualify the headline numbers.
  2. [Appx. C] Typo: 'In In the experiment' should be 'In the experiment'.
  3. [Sec. 4.1.1 / Eq. (3)] Reference [10] is a web tutorial for percentile transforms; please cite a standard statistical reference instead.
  4. [Alg. 2 / Sec. 4.2] The phrase 'Add ghost node for sequences in T_u^w' is ambiguous: ghost nodes are described as representing missing events, not missing sequences. Please align the terminology.
  5. [References] References [8] and [51] appear to be the same LightPath paper; please deduplicate.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central derivation is self-contained. Minor self-citations are not load-bearing, and the ghost-node under-specification is a reproducibility gap, not a circular reduction.

full rationale

The paper's anomaly scores reuse the model's own pretraining objectives (node reconstruction and link prediction), but this is the standard unsupervised autoencoder-style anomaly detection paradigm: the model is trained only on normal mobility data, and anomalies are held-out test events. The observed-link score f_AS(e)_obs-link and the absence score f_AS(e)_not-obs-link are computed from the model's predicted link likelihood at inference; they are not fitted to anomaly labels and therefore are not predictions that reduce to their own inputs by construction. The absence-anomaly score is a legitimate counterfactual: if the model assigns high likelihood to an unobserved co-occurrence, the event is scored anomalous. The simulated anomaly injection matches the paper's explicit definition of co-occurrence (Section 3.2) and its two anomaly types, which is a benchmark-construction concern rather than a circular derivation. The paper does contain several self-citations in related work and architectural inspiration (e.g., refs. [35], [42], [52], [56]), but none is load-bearing for the central claim: no uniqueness theorem or model component is justified solely by these self-citations, and the core methodology is independently specified. The main genuine weakness is the under-specified 'ghost node' mechanism in Eq. (12) and Algorithm 2 line 1: the paper never states how many ghost nodes are added, what features they receive, or how their embeddings enter the cross-people attention. This is an omitted implementation detail that harms reproducibility, but it is not circularity because no equation or fitted parameter is being smuggled in as a prediction. Overall, no load-bearing circular step was found; the score reflects only the presence of minor, non-controlling self-citations.

Assumptions & free parameters 8 free parameters · 4 assumptions · 1 invented entities

The central claim depends on several hyperparameters and domain assumptions. The most significant is the co-occurrence definition, which is also used to generate the evaluation anomalies. The ghost node is a practical modeling device without independent validation.

free parameters (8)
  • mask ratio = 0.05 (searched in [0.05, 0.3])
    Masked pre-training ratio, selected by hyperparameter search.
  • embedding size D = searched in [32, 64, 128]
    Model capacity hyperparameter, tuned for best performance.
  • number of negative edges = 5
    Fixed number of negative samples per positive link in contrastive loss.
  • number of TSA layers M = 1
    Stacking depth of the two-stage attention, chosen by search.
  • attention heads H = 4
    Number of heads in cross-people attention.
  • time window w = 3 days
    Length of input event sequences.
  • co-occurrence distance threshold delta = 40 meters
    Defines interaction between events; central to the collective behavior definition.
  • link loss weight lambda = 0.01
    Balancing node and link reconstruction losses.
assumptions (4)
  • domain assumption Stay events are defined as GPS locations stationary for at least five minutes.
    Section 3.1. This threshold determines what counts as an event; may not hold for all datasets.
  • domain assumption Two events co-occur if spatial distance < 40m and time intervals overlap.
    Section 3.2 and Appendix B. This defines collective behavior; if co-occurrence is not a proxy for interaction, the model's semantics change.
  • domain assumption Related individuals are those with co-occurrence in the window or frequent meetings (at least twice, total duration > 2h).
    Section 3.2. The definition of the collective event sequence is based on these thresholds.
  • domain assumption The simulated datasets faithfully represent realistic human mobility and collective behavior.
    Section 5.1.1. The evaluation trusts the simulation and the injected anomalies as ground truth.
invented entities (1)
  • Ghost node
    purpose: Represents a potential missing neighbor event to compute the absence-anomaly score.
    Introduced in Section 4.2 and Algorithm 2. Its feature construction is not specified, and no external evidence validates that it correctly represents possible missing events.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoBAD: Modeling Collective Behaviors for Human Mobility Anomaly Detection." pith.science (2026). https://pith.science/paper/PJPLHGF2

@misc{pith2026250814088,
  author       = {Pith},
  title        = {Pith review of: CoBAD: Modeling Collective Behaviors for Human Mobility Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJPLHGF2}},
  note         = {Machine review of arXiv:2508.14088}
}
read the original abstract

Detecting anomalies in human mobility is essential for applications such as public safety and urban planning. While traditional anomaly detection methods primarily focus on individual movement patterns (e.g., a child should stay at home at night), collective anomaly detection aims to identify irregularities in collective mobility behaviors across individuals (e.g., a child is at home alone while the parents are elsewhere) and remains an underexplored challenge. Unlike individual anomalies, collective anomalies require modeling spatiotemporal dependencies between individuals, introducing additional complexity. To address this gap, we propose CoBAD, a novel model designed to capture Collective Behaviors for human mobility Anomaly Detection. We first formulate the problem as unsupervised learning over Collective Event Sequences (CES) with a co-occurrence event graph, where CES represents the event sequences of related individuals. CoBAD then employs a two-stage attention mechanism to model both the individual mobility patterns and the interactions across multiple individuals. Pre-trained on large-scale collective behavior data through masked event and link reconstruction tasks, CoBAD is able to detect two types of collective anomalies: unexpected co-occurrence anomalies and absence anomalies, the latter of which has been largely overlooked in prior work. Extensive experiments on large-scale mobility datasets demonstrate that CoBAD significantly outperforms existing anomaly detection baselines, achieving an improvement of 13%-18% in AUCROC and 19%-70% in AUCPR. All source code is available at https://github.com/wenhaomin/CoBAD.

Figures

Figures reproduced from arXiv: 2508.14088 by the authors.

Figure 1
Figure 1. Illustration of Individual and Collective Spatio [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of Collective Event Sequences (CES) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. CoBAD model architecture, which learns collective human mobility patterns in a self-supervised fashion via masked pretraining. CoBAD employs a two-stage attention layer with (1) a cross-time attention module to capture the spatiotemporal dependencies in an event sequence of each single individual and (2) a cross-people attention module based on the graph transformer, to capture the collective interactions or relatio… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Negative Edge Sampling. For each positive edge [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of two collective anomaly types. Obs: [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: CoBAD can detect collective anomalies of various types. Ablation study of CoBAD variants on different anom￾alies: adding CoBAD-obs-link loss to node reconstruction loss boosts performance significantly on unexpected occur￾rence and synthetic coordination anomalies, whi…
Figure 7
Figure 7. Figure 7: (best viewed in color) Case study: A detected “absence anomaly” (i.e., high score in the missing link). Visualization [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: (best viewed in color) Case study: A detected "unexpected occurrence anomaly" (i.e., high score in the observed [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 10
Figure 10. Figure 10: CoBAD uses memory efficiently, proportionate to input size. Left: GPU Memory Usage per batch w.r.t 𝐿 when 𝑁=5, right: GPU Memory Usage per batch w.r.t 𝑁 when 𝐿=10. Common setting: batch size = 128, 𝐷=32 [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 9
Figure 9. Figure 9: CoBAD scales near-linearly with input size. Left: training time per sample w.r.t 𝐿 when 𝑁=5, right: training time per sample w.r.t 𝑁 when 𝐿=10. Common setting: batch size = 128, 𝐷=32. 0 100 200 300 L (Sequence Length) 5000 10000 15000 20000 GPU Memory Usage (MB) 0 20 4…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 48 canonical work pages

  1. [1]

    Zain Ul Abideen, Heli Sun, Zhou Yang, Rana Zeeshan Ahmad, Adnan Iftekhar, and Amir Ali. 2021. Deep Wide Spatial-Temporal Based Transformer Networks Mod- eling for the Next Destination According to the Taxi Driver Behavior Prediction. Applied Sciences 11, 1 (2021)

    work page 2021

  2. [2]

    Peter Adey. 2004. Surveillance at the airport: surveilling mobility/mobilising surveillance. Environment and Planning A 36, 8 (2004), 1365–1380

    work page 2004

  3. [3]

    Joseph B Bak-Coleman, Mark Alfano, Wolfram Barfuss, Carl T Bergstrom, Miguel A Centeno, Iain D Couzin, Jonathan F Donges, Mirta Galesic, Andrew S Gersick, Jennifer Jacquet, et al. 2021. Stewardship of global collective behavior. Proceedings of the National Academy of Sciences 118, 27 (2021), e2025764118

    work page 2021

  4. [4]

    Hugo Barbosa, Marc Barthelemy, Gourab Ghoshal, Charlotte R James, Maxime Lenormand, Thomas Louail, Ronaldo Menezes, José J Ramasco, Filippo Simini, and Marcello Tomasini. 2018. Human mobility: Models and applications. Physics Reports 734 (2018), 1–74

    work page 2018

  5. [5]

    Claudio Castellano, Santo Fortunato, and Vittorio Loreto. 2009. Statistical physics of social dynamics. Reviews of modern physics 81, 2 (2009), 591–646

    work page 2009

  6. [6]

    Buru Chang, Yonggyu Park, Donghyeon Park, Seongsoon Kim, and Jaewoo Kang. 2018. Content-aware hierarchical point-of-interest embedding model for successive poi recommendation.. In IJCAI, Vol. 20. 3301–3307

    work page 2018

  7. [7]

    Meng Chen, Yang Liu, and Xiaohui Yu. 2014. NLPMM: A Next Location Predictor with Markov Modeling. In Advances in Knowledge Discovery and Data Mining , Vincent S. Tseng, Tu Bao Ho, Zhi-Hua Zhou, Arbee L. P. Chen, and Hung-Yu Kao (Eds.). Springer International Publishing, Cham, 186–197

    work page 2014

  8. [8]

    Yile Chen, Xiucheng Li, Gao Cong, Zhifeng Bao, Cheng Long, Yiding Liu, Arun Ku- mar Chandran, and Richard Ellison. 2021. Robust road network representation learning: When traffic patterns meet traveling semantics. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management . 211–220

    work page 2021

Show all 60 references
  1. [9]

    Lyu, and Irwin King

    Chen Cheng, Haiqin Yang, Michael R. Lyu, and Irwin King. 2013. Where you like to go next: successive point-of-interest recommendation. In Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence (Beijing, China) (IJCAI ’13). AAAI Press, 2605–2611

    2013

  2. [10]

    Cuemath. n.d.. Percentile Formula - How to Calculate Percentile? https://www. cuemath.com/percentile-formula/ Accessed: 2025-01-25

    2025

  3. [11]

    D. J. Daley and David Vere-Jones. 2008. An Introduction to the Theory of Point Processes: Volume II (2nd ed.). Springer

    2008

  4. [12]

    Nan Du, Hanjun Dai, Rakshit Trivedi, Utkarsh Upadhyay, Manuel Gomez- Rodriguez, and Le Song. 2016. Recurrent marked temporal point processes: Embedding event history to vector. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data minin...

    2016

  5. [13]

    Vijay Prakash Dwivedi and Xavier Bresson. 2020. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699 (2020)

    2020 arXiv

  6. [14]

    Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng Jin. 2018. DeepMove: Predicting Human Mobility with Attentional Recurrent Networks. In Proceedings of the 2018 World Wide Web Conference (Lyon, France) (WWW ’18). International World Wide Web Conferen...

    2018

  7. [15]

    Yuye Feng, Wei Zhang, Yao Fu, Weihao Jiang, Jiang Zhu, and Wenqi Ren. 2024. SensitiveHUE: Multivariate Time Series Anomaly Detection by Enhancing the Sensitivity to Normal Patterns. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24...

    2024

  8. [16]

    Sébastien Gambs, Marc-Olivier Killijian, and Miguel Núñez del Prado Cortez

  9. [17]

    Qiang Gao, Xiaohan Wang, Chaoran Liu, Goce Trajcevski, Li Huang, and Fan Zhou. 2023. Open anomalous trajectory recognition via probabilistic metric learning. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI ’23). Article 233, 9 pages

    2023

  10. [18]

    Qiang Gao, Fan Zhou, Kunpeng Zhang, Goce Trajcevski, Xucheng Luo, and Fengli Zhang. 2017. Identifying human mobility via trajectory embeddings. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI’17). 1689–1695

    2017

  11. [19]

    Shengnan Guo, Youfang Lin, Ning Feng, Chao Song, and Huaiyu Wan. 2019. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI conference on artificial intelligence , Vol. 33. 922–929

    2019

  12. [20]

    Sheng, and Jian Yang

    Venus Haghighi, Behnaz Soltani, Nasrin Shabani, Jia Wu, Yang Zhang, Lina Yao, Quan Z. Sheng, and Jian Yang. 2024. TROPICAL: Transformer-Based Hyper- graph Learning for Camouflaged Fraudster Detection. In 2024 IEEE International Conference on Data Mining (ICDM) . 121–130. doi:1...

    2024

  13. [21]

    Alan G. Hawkes. 1971. Spectra of some self-exciting and mutually exciting point processes. Biometrika 58 (1971), 83–90

    1971

  14. [22]

    Zhixiang He, Chi-Yin Chow, and Jia-Dong Zhang. 2019. STCNN: A spatio- temporal convolutional neural network for long-term traffic prediction. In 2019 20th IEEE international conference on mobile data management (MDM) . IEEE, 226–233

    2019

  15. [23]

    Pablo Jensen. 2019. The politics of physicists’ social models. Comptes Rendus. Physique 20, 4 (2019), 380–386

    2019

  16. [24]

    Jiawei Jiang, Dayan Pan, Houxing Ren, Xiaohan Jiang, Chao Li, and Jingyuan Wang. 2023. Self-supervised trajectory representation learning with temporal regularities and travel semantics. In 2023 IEEE 39th international conference on data engineering (ICDE). IEEE, 843–855

    2023

  17. [25]

    Maik Larooij and Petter Törnberg. 2025. Do Large Language Models Solve the Problems of Agent-Based Modeling? A Critical Review of Generative Social Simulations. arXiv preprint arXiv:2504.03274 (2025)

    2025 arXiv

  18. [26]

    Chaoneng Li, Guanwen Feng, Yunan Li, Ruyi Liu, Qiguang Miao, and Liang Chang. 2024. DiffTAD: Denoising diffusion probabilistic models for vehicle trajectory anomaly detection. Knowledge-Based Systems 286 (2024), 111387

    2024

  19. [27]

    Yan Lin, Huaiyu Wan, Shengnan Guo, and Youfang Lin. 2021. Pre-training Context and Time Aware Location Embeddings from Spatial-Temporal Trajectories for User Next Location Prediction. Proceedings of the AAAI Conference on Artificial Intelligence 35 (May 2021), 4241–4248

    2021

  20. [28]

    Yiding Liu, Kaiqi Zhao, Gao Cong, and Zhifeng Bao. 2020. Online Anomalous Trajectory Detection with Deep Generative Sequence Modeling. In 2020 IEEE 36th International Conference on Data Engineering (ICDE) . 949–960

    2020

  21. [29]

    Gary T Marx and Doug McAdam. 1994. Collective behavior and social movements: Process and structure. (No Title) (1994)

    1994

  22. [30]

    Sandro Meloni, Nicola Perra, Alex Arenas, Sergio Gómez, Yamir Moreno, and Alessandro Vespignani. 2011. Modeling human mobility responses to the large- scale spreading of infectious diseases. Scientific reports 1, 1 (2011), 62

    2011

  23. [31]

    Anna Monreale, Fabio Pinelli, Roberto Trasarti, and Fosca Giannotti. 2009. WhereNext: a location predictor on trajectory pattern mining. In Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’09) . 637–646

    2009

  24. [32]

    Yosihiko Ogata. 1998. Space-Time Point-Process Models for Earthquake Occur- rences. Annals of the Institute of Statistical Mathematics 50 (1998), 379–402

    1998

  25. [33]

    Elin Palm. 2013. Rights that trump: Surveillance-based migration governance and a substantial right to mobility. Journal of information, communication and ethics in society 11, 4 (2013), 196–209

    2013

  26. [34]

    Tim Poštuvan, Claas Grohnfeldt, Michele Russo, and Giulio Lovisotto. 2024. Learning-Based Link Anomaly Detection in Continuous-Time Dynamic Graphs. arXiv preprint arXiv:2405.18050 (2024)

    2024 arXiv

  27. [35]

    Shebuti Rayana and Leman Akoglu. 2015. Collective opinion spam detection: Bridging review networks and metadata. In Proceedings of the 21th acm sigkdd international conference on knowledge discovery and data mining . 985–994

    2015

  28. [36]

    Emanuele Rossi, Ben Chamberlain, Fabrizio Frasca, Davide Eynard, Federico Monti, and Michael Bronstein. 2020. Temporal graph networks for deep learning on dynamic graphs. arXiv preprint arXiv:2006.10637 (2020)

    2020 arXiv

  29. [37]

    Saeedreza Shehnepoor, Roberto Togneri, Wei Liu, and Mohammed Bennamoun

  30. [38]

    Li Song, Ruijia Wang, Ding Xiao, Xiaotian Han, Yanan Cai, and Chuan Shi. 2018. Anomalous trajectory detection using recurrent neural network. In Advanced Data Mining and Applications: 14th International Conference, ADMA 2018, Nanjing, China, November 16–18, 2018, Proceedings 1...

    2018

  31. [39]

    Xuan Song, Hiroshi Kanasugi, and Ryosuke Shibasaki. 2016. Deeptransport: prediction and simulation of human mobility and transportation mode at a citywide level. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI’16). 2618–2624

    2016

  32. [40]

    Chris Stanford, Suman Adari, Xishun Liao, Yueshuai He, Qinhua Jiang, Chenchen Kuai, Jiaqi Ma, Emmanuel Tung, Yinlong Qian, Lingyi Zhao, Zihao Zhou, Zeeshan Rasheed, and Khurram Shafique. 2024. NUMOSIM: A Synthetic Mobility Dataset with Anomaly Detection Benchmarks. arXiv:2409....

    2024 arXiv

  33. [41]

    Glen Van Brummelen. 2017. Heavenly Mathematics: The Forgotten Art of Spherical Trigonometry. Princeton University Press. https://doi.org/10.23943/princeton/ 9780691175997.001.0001 ACM SIGSPATIAL’25, November 3–6, 2025, Minneapolis, MN, USA Haomin Wen et al

    2017 doi

  34. [42]

    Véronique Van Vlasselaer, Leman Akoglu, Tina Eliassi-Rad, Monique Snoeck, and Bart Baesens. 2015. Guilt-by-constellation: Fraud detection by suspicious clique memberships. In 2015 48th Hawaii International Conference on System Sciences . IEEE, 918–927

    2015

  35. [43]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL]

    2023 arXiv

  36. [44]

    Huaiyu Wan, Yan Lin, Shengnan Guo, and Youfang Lin. 2021. Pre-training time- aware location embeddings from spatial-temporal trajectories. IEEE Transactions on Knowledge and Data Engineering 34, 11 (2021), 5510–5523

    2021

  37. [45]

    Weiying Wang and Toshihiro Osaragi. 2024. Learning Daily Human Mobility with a Transformer-Based Model. ISPRS International Journal of Geo-Information 13, 2 (2024)

    2024

  38. [46]

    Lixia Wu, Haomin Wen, Haoyuan Hu, Xiaowei Mao, Yutong Xia, Ergang Shan, Jianbin Zheng, Junhong Lou, Yuxuan Liang, Liuqing Yang, et al. 2024. LaDe: The First Comprehensive Last-mile Express Dataset from Industry. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Dis...

    2024

  39. [47]

    Xian Wu, Chao Huang, Chuxu Zhang, and Nitesh V. Chawla. 2020. Hierarchically Structured Transformer Networks for Fine-Grained Spatial Event Forecasting. In Proceedings of The Web Conference 2020 (WWW ’20) . 2320–2330

    2020

  40. [48]

    Da Xu, Chuanwei Ruan, Evren Korpeoglu, Sushant Kumar, and Kannan Achan

  41. [49]

    Hao Xue, Flora Salim, Yongli Ren, and Nuria Oliver. 2021. MobTCast: Leveraging Auxiliary Trajectory Forecasting for Human Mobility Prediction. In Advances in Neural Information Processing Systems , Vol. 34. 30380–30391

    2021

  42. [50]

    Sean Bin Yang, Jilin Hu, Chenjuan Guo, Bin Yang, and Christian S. Jensen. 2023. LightPath: Lightweight and Scalable Path Representation Learning. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). 2999–3010

    2023

  43. [51]

    Sean Bin Yang, Jilin Hu, Chenjuan Guo, Bin Yang, and Christian S Jensen. 2023. Lightpath: Lightweight and scalable path representation learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2999–3010

    2023

  44. [52]

    Junting Ye and Leman Akoglu. 2015. Discovering opinion spammer groups by network footprints. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2015, Porto, Portugal, September 7-11, 2015, Proceedings, Part I 15 . Springer, 267–282

    2015

  45. [53]

    Bing Yu, Haoteng Yin, and Zhanxing Zhu. [n. d.]. Spatio-Temporal Graph Convo- lutional Networks: A Deep Learning Framework for Traffic Forecasting. ([n. d.])

  46. [54]

    Daqing Zhang, Nan Li, Zhi-Hua Zhou, Chao Chen, Lin Sun, and Shijian Li. 2011. iBAT: detecting anomalous taxi trajectories from GPS traces. InProceedings of the 13th International Conference on Ubiquitous Computing (Beijing, China) (UbiComp ’11). 99–108

    2011

  47. [55]

    Ling Zhao, Yujiao Song, Chao Zhang, Yu Liu, Pu Wang, Tao Lin, Min Deng, and Haifeng Li. 2019. T-GCN: A temporal graph convolutional network for traffic prediction. IEEE transactions on intelligent transportation systems 21, 9 (2019), 3848–3858

    2019

  48. [56]

    Yiji Zhao, Youfang Lin, Haomin Wen, Tonglong Wei, Xiyuan Jin, and Huaiyu Wan. 2022. Spatial-temporal position-aware graph convolution networks for traffic flow forecasting. IEEE Transactions on Intelligent Transportation Systems 24, 8 (2022), 8650–8666

    2022

  49. [57]

    Yu Zheng, Lizhu Zhang, Xing Xie, and Wei-Ying Ma. 2009. Mining interesting locations and travel sequences from GPS trajectories. In WWW. 791–800. A Dataset Details The summary statistics of the datasets are in Table 4. B Model Configurations For all deep models, we search the ...

    2009

  50. [2012]

    In Proceedings of the First Workshop on Measurement, Privacy, and Mobility

    Next place prediction using mobility Markov chains. In Proceedings of the First Workshop on Measurement, Privacy, and Mobility . Article 3, 6 pages

  51. [2020]

    arXiv preprint arXiv:2002.07962 (2020)

    Inductive representation learning on temporal graphs. arXiv preprint arXiv:2002.07962 (2020)

    2002 arXiv

  52. [2023]

    IEEE Transactions on Neural Networks and Learning Systems 34, 8 (2023), 4153–4166

    HIN-RNN: A Graph Representation Learning Neural Network for Fraudster Group Detection With No Handcrafted Features. IEEE Transactions on Neural Networks and Learning Systems 34, 8 (2023), 4153–4166. doi:10.1109/TNNLS.2021. 3123876

    2023 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.