Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Secrets Everywhere: Auditing Memorization in Mobility Prediction Models

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that mobility prediction models memorize large shares of user trajectories—up to 85% in some settings—and that the degree of memorization predicts how easily an attacker can reconstruct those trajectories from partial know

desk verdict Useful first trajectory-level memorization audit for mobility, but medoid-based reference sets conflate memorization with typicality, so the headline numbers need a control. read the letter →

arxiv 2608.02052 v1 pith:2XOEJ6CA submitted 2026-08-03 cs.LG

Anne Josiane Kouam , Hristo Boyadzhiev , Konrad Rieck This is my paper
classification cs.LG
keywords memorizationmobilitypredictionprivacyauditingtrajectoryextractionexposuremetriclocationreferencesetsdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to establish that memorization is a pervasive and measurable privacy risk in mobility prediction models, not a rare anomaly. It argues that the standard exposure metric used for language models undercounts this risk because mobility secrets are ordinary, inlier trajectories rather than outliers. To fix this, the authors introduce a framework that measures how much more likely a model is to prefer each training trajectory over a set of behaviorally similar reference trajectories, at three granularities: individual locations, anchor pairs, and short sub-trajectories. Their experiments across three public datasets and six model architectures show strong memorization signals for large fractions of trajectories, with the highest-memorized trajectories being the ones easiest to recover in simulated extraction attacks. If correct, this means privacy auditing of mobility models cannot rely on existing rarity-based metrics and needs to treat every user trajectory as a potential leak.

What carries the argument

The load-bearing apparatus is the behavioral abstraction, a mapping from a raw trajectory to a representation of the mobility pattern the model is expected to generalize from (motion vectors for location memorization, home/work anchor plus radius for anchor-pair memorization, a masked slot for segment-level memorization). Around each abstraction, the framework builds a reference set: for segment-level risks, by substituting, shuffling, or making stationary plausible short windows from the same user; for location and anchor risks, by clustering the abstractions of all trajectories and taking the cluster medoid as the training sample and the other cluster members as references. Memorization is

What would settle it

Train a model on trajectories that are deliberately not the medoid of their cluster and compute the same metrics; if memorization scores drop sharply compared to the medoid-based training, the reported memorization is an artifact of selecting the medoid as the training sample. Alternatively, take a held-out trajectory, make it the medoid of a cluster, and check whether it receives similarly high scores despite never being seen in training.

Watch

Extended reading notes

Core claim

The central discovery is that mobility prediction models systematically assign abnormally high likelihood to their training trajectories compared with held-out trajectories that share the same behavioral abstraction. Three complementary metrics—adapted exposure, exposure preference, and exposure magnitude—quantify this preference per trajectory. Across datasets, between 87% and over 99% of training trajectories show positive memorization magnitude, and in the most routinized dataset 99.4% of trajectories sit among the most preferred alternatives. Memorization is strongest for regular, low-entropy users, appears even in simple Markov and small-capacity neural models, correlates with reduced u

Load-bearing premise

The central assumption is that the most representative trajectory of a behavioral cluster—the medoid—is an appropriate stand-in for any training trajectory; if a model prefers the medoid simply because it is typical of the cluster's behavior, the measured 'memorization' is inflated and the metrics would not distinguish learning a pattern from memorizing the data.

Editorial extensions

If this is right

  • Privacy audits of mobility prediction models should treat every trajectory as a potential secret and report distribution-level memorization scores, not a single global number.
  • The new metrics allow comparing memorization across architectures, datasets, and user groups, enabling targeted interventions such as stronger regularization for routinized users.
  • Because memorization magnitude correlates with extraction success, the metrics can serve as a pre-deployment early-warning tool for which users are most exposed to inference attacks.
  • The observed negative correlation between memorization and utility suggests a trade-off frontier: models that generalize better also tend to leak less of their training data in these settings.

Reading between the lines

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

  • The medoid-selection step may confound memorization with representativeness: a model could prefer the most typical trajectory of a behavioral cluster even without memorization, so the high scores reported here might partly reflect cluster structure rather than data retention.
  • The framework's three granularities suggest an extension to cross-user memorization, where population-level popular locations are learned more strongly; the authors note this as future work, but one could test it by measuring location-level memorization conditional on location popularity.
  • The same reference-set methodology could transfer to other sequential prediction domains with inlier secrets, such as next-event prediction in healthcare or financial transactions, where privacy leaks are equally pervasive.
  • If the utility–memorization correlation holds generally, it implies a design tension for mobility services: improving prediction accuracy by training on real human traces may inevitably increase the exposure of those traces, strengthening the case for differential privacy or on-device training.
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 / 5 minor

Summary. The paper proposes a framework for auditing memorization in mobility prediction models, introducing three trajectory-level metrics—exposure, exposure preference, and exposure magnitude—computed against user-grounded reference sets. It defines three abstractions (location, anchor-pair, segment) and constructs reference sets via clustering medoids (for location and anchor-pair) or via perturbing segments of the same user (for segment-level). Evaluation on three mobility datasets and seven model variants reports pervasive memorization, correlations with user regularity, and a claimed link between memorization and extractability. The paper also compares its metrics with the original Carlini et al. exposure and discusses implications for privacy auditing.

Significance. If the central claim holds, this is a valuable first systematic audit of memorization in mobility prediction: it addresses the inlier nature of mobility secrets, proposes metrics that go beyond synthetic canaries, and provides an openly released artifact and an extensive multi-dataset, multi-model evaluation. The honest appendix reporting weaker or inconsistent effects in several settings is a credit. However, the medoid-based training design leaves the headline claim vulnerable to a typicality confound that is not addressed by any control experiment, so the pervasive-memorization conclusion is not yet established.

major comments (3)
  1. [Section 4.2.2, Definition 4.3, Algorithm 1] The approximate reference-set construction selects the cluster medoid as the training trajectory. Since the medoid is by definition the most representative trajectory in the abstraction space, a model that has learned the cluster's behavioral distribution without memorizing any specific sequence will assign it higher likelihood than peripheral reference trajectories. Thus preference(𝑇𝑢)=rank/|R| and magnitude(𝑇𝑢)=L(𝑇𝑢)−E[L(𝑇′)] conflate memorization with typicality. The spatial-dispersion analysis (Figure 10) addresses geographic spread, not likelihood typicality. The RQ2 headline ('up to 85% of training trajectories show strong memorization signals') is therefore not established without a control experiment, e.g., training on a random cluster member or on held-out medoids and showing that the high scores disappear.
  2. [Section 6.4, Figure 7, Appendix D.3] The abstract and introduction claim a 'clear correlation' between higher memorization scores and fewer extraction attempts (RQ4). The appendix, however, reports that prefix-based separation is 'less pronounced' in Shenzhen Urban and that anchor-based extractability is 'substantially overlapping' or 'nearly indistinguishable' in other datasets. The main text should either report these weaker effects quantitatively with appropriate statistical tests or temper the claim. As written, the central extractability conclusion overstates the evidence.
  3. [Section 6.3, Figure 5] The paper states that there is 'no correlation' between the proposed trajectory-level metrics and Carlini et al.'s model-level exposure. But Figure 5 plots one Carlini score per model as a vertical marker against a kernel density; no correlation coefficient or multi-model quantitative comparison is provided. With only seven model variants, a visual comparison of a scalar to a distribution does not support a claim of 'no correlation.' Either compute a model-level correlation across the variants or soften the claim to 'no consistent alignment.'
minor comments (5)
  1. [Section 6.3] Typo: 'exposure exposure tends to be higher' should read 'exposure tends to be higher.'
  2. [Figure 13 caption] Typo: 'trainined' should be 'trained.'
  3. [Figure 15 caption and Section 5.1] The dataset is referred to as 'Shanghai Telecom' in the text but 'Shanghai Kaggle' in a figure caption. Please unify the naming.
  4. [Algorithm 1] The algorithm is described in the main text but referenced as 'Appendix 1.' Also, the pseudocode uses 𝜏 as a hard threshold, while the text says the effective threshold is implicit; clarify the relationship.
  5. [Section 5.4] The value of 𝑘min=100 is justified as a 'rule of thumb,' but no sensitivity analysis is provided. Since the metrics depend on reference-set size, a short robustness check (e.g., 𝑘min ∈ {50,200}) would strengthen the claims.

Circularity Check

2 steps flagged · score 6.0 of 10

Two central results reduce to the construction: medoid reference sets make 'memorization' track typicality, and extractability re-ranks the same likelihoods used to define memorization.

  1. self definitional [Section 4.3 (metric definitions) and Section 6.4 / Appendix D (Eq. 2, Extractability_prefix)]
    "While the attack objectives differ, both settings rely on a common underlying signal, namely, the model’s tendency to assign disproportionately high likelihood to memorized data. ... Extractability is measured as the logarithm of the number of greedy attempts required to retrieve the ground-truth suffix within the decoding tree."

    Memorization magnitude is L(Tu) - E[L(T')] and preference is rank(L(Tu))/|R(Tu)|, both monotone functions of the model likelihood L. The prefix extractability score is defined as log10(rank_suffix), where rank_suffix is the rank of the true suffix among beam-search candidates scored by the same L. Therefore the reported correlation (higher magnitude -> fewer attempts) is a mathematical consequence of both metrics using the same likelihood signal, not an independent empirical test. The paper's sentence that both settings rely on a common underlying signal concedes this, so the RQ4 'prediction' reduces to the definition of the metrics.

  2. self definitional [Section 4.2.2 / Definition 4.3 / Algorithm 1, applied in Section 5.4]
    "From each cluster, we then select a representative trajectory Tu to serve as the training instance, while all other trajectories in the same cluster form its reference set. ... Select medoid trajectory Tu in C as training sample"

    By Definition 4.3 and Algorithm 1, the training trajectory Tu is chosen as the cluster medoid — the most representative trajectory of the behavioral cluster. The reference set R(Tu) contains the remaining, by construction more peripheral cluster members. A non-memorizing model that has learned the cluster distribution will assign higher likelihood to the medoid than to boundary trajectories, so preference(Tu)=rank(L(Tu))/|R(Tu)| and magnitude(Tu)=L(Tu)-E[L(T')] will register positive 'memorization' even absent memorization. The premise in Section 4.2 that a generalizing model should assign comparable likelihoods to the original and reference trajectories is violated by this medoid selection, making the pervasive-memorization finding partly an artifact of the construction. No control traini

full rationale

The paper's framework is not circular in the sense of fitting parameters to force the outcome: the exposure/preference/magnitude metrics are computed directly from model likelihoods and reference sets, and no self-citation chain is load-bearing. The central empirical claims, however, contain two construction-driven reductions. First, the approximate reference-set construction selects the cluster medoid as the training trajectory; since a medoid is by definition the most typical member, the memorization metrics conflate memorization with typicality, so the headline 'up to 85% of training trajectories show strong memorization signals' is not established independently of this selection. Second, the extractability validation re-ranks trajectories using the same model likelihoods that define the memorization metrics, so the RQ4 correlation is tautological rather than an independent demonstration of extraction risk. These circular components affect the paper's two most prominent conclusions (pervasive memorization and extractability), while the descriptive analyses of user regularity, model differences, and the comparison to Carlini et al.'s exposure retain independent empirical content. Score 6 reflects partial circularity: some predictions reduce by construction, but not the entire derivation.

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

The framework introduces no new physical or mathematical entities. Its load-bearing components are the behavioral abstractions and reference-set construction rules, whose free parameters (kmin, cluster count, training size, implicit tau) are hand-chosen and affect the measured memorization levels. The central untested premise is that medoid-vs-reference likelihood differences measure memorization rather than typicality.

free parameters (4)
  • kmin (minimum reference set size) = 100
    Chosen as a 'rule of thumb' in Section 5.4 to balance stability and behavioral coherence; directly controls reference set difficulty and thus memorization scores.
  • Number of clusters = 2000
    Fixed to equal the training set size in Section 5.4; determines which trajectories become training medoids and the size/coherence of reference sets.
  • Training set size = 2000 trajectories
    Fixed in Section 5.4 for all experiments; smaller training sets may increase memorization, so results depend on this choice.
  • Implicit enrichment threshold tau = dataset-dependent (e.g., 1.67 for Shenzhen location, 45.31 for YJMob anchor)
    Described in Section 5.4 as an implicit threshold: trajectories are added to reference sets in increasing distance order until kmin is reached. This makes reference-set difficulty vary by dataset and abstraction.
assumptions (4)
  • domain assumption A non-memorizing model should assign comparable likelihood to any trajectory satisfying the same behavioral abstraction.
    This is the core premise behind all three metrics, stated in Sections 4.1 and 4.2. If the abstraction is too coarse, the model may legitimately prefer the training trajectory because it is more typical, not because it is memorized.
  • domain assumption The cluster medoid is a representative training trajectory and the other cluster members are behaviorally equivalent references.
    Algorithm 1 selects the medoid as the training sample. References are peripheral by construction, so a higher likelihood for the medoid may reflect centrality rather than memorization. The paper does not test this confound.
  • domain assumption Negative log-perplexity ranking is a valid proxy for memorization in mobility models.
    Adopted from Carlini et al. (Section 4.3), this assumes that likelihood gaps between training and reference trajectories correspond to retention of specific training data rather than to a model's learned typicality distribution.
  • standard math Standard clustering, beam search, and likelihood computation behave as assumed.
    k-means clustering, Euclidean distances in abstraction space, and beam search decoding are used as off-the-shelf tools without modification or formal verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Secrets Everywhere: Auditing Memorization in Mobility Prediction Models." pith.science (2026). https://pith.science/paper/2XOEJ6CA

@misc{pith2026260802052,
  author       = {Pith},
  title        = {Pith review of: Secrets Everywhere: Auditing Memorization in Mobility Prediction Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XOEJ6CA}},
  note         = {Machine review of arXiv:2608.02052}
}
read the original abstract

Human mobility prediction models, which forecast the next location in a user's trajectory, are increasingly deployed in urban analytics, navigation, and personalized services. Yet, little is known about their potential to memorize and expose sensitive user trajectories from training data. While memorization has been extensively studied in language models, mobility prediction poses unique challenges: training sequences encode human behavior at various spatial and temporal scales, creating privacy risks at different granularities. In this paper, we conduct the first systematic audit of memorization in mobility prediction models. While prior work has shown that privacy leaks can arise from such models, we systematically assess and quantify memorization risks at scale. We identify key challenges, including the lack of a randomness space, the multi-scale structure of trajectories, and user-specific behavioral diversity. To address these challenges, we introduce a framework to quantify mobility memorization at different levels of granularity: individual locations, anchor pairs, and subtrajectory segments. We also develop user-grounded reference sets to assess how likely a model is to prefer training data over realistic alternatives. Our evaluation across multiple models and datasets reveals pervasive memorization patterns that correlate with user regularity and increase the risk of data extraction at inference time. Our findings call for mandatory privacy auditing in mobility prediction models.

Figures

Figures reproduced from arXiv: 2608.02052 by the authors.

Figure 1
Figure 1. Overview of our memorization auditing framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Cumulative distribution of mobility memorization [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Mobility characteristics and memorization patterns for [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Memorization behavior across risks, transformations, and sub-trajectories ( [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Exposure distributions across models on the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Utility (Top-1 Accuracy) vs. Memorization (Mag [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 8
Figure 8. Figure 8: Distribution of mobility diversity across datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: CDF of reference-set sizes across datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: CDF of average spatial dispersion within reference [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Training dynamics for the LSTM-simple model across all datasets: Training and validation loss decrease steadily and [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Mobility characteristics and memorization patterns: Exposure preference for location memorization on the model [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Memorization behavior across risks, transformations, and sub-trajectories of the following models trainined on [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Comparison of memorization exposure between [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Exposure distributions across model architectures, with Carlini et al [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]
Figure 16
Figure 16. Figure 16: Memorization–utility relationship (magnitude vs. Top-1 Accuracy) across three datasets for the following models: [PITH_FULL_IMAGE:figures/full_fig_p020_16.png]
Figure 17
Figure 17. Figure 17: Extractability versus memorization signals for LSTM-simple across two datasets (a)-(b) Shenzhen Urban and (c)-(d) [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 7 canonical work pages

  1. [1]

    Amichi, A

    L. Amichi, A. C. Viana, M. Crovella, and A. A.F. Loureiro. 2020. Understanding individuals’ proclivity for novelty seeking. InACM SIGSPATIAL. doi:10.1145/ 3397536.3422248

    arXiv 2020

  2. [2]

    Shenoy, and David D

    Akanksha Atrey, Prashant J. Shenoy, and David D. Jensen. 2021. Preserving Privacy in Personalized Models for Distributed Mobile Services. In41st IEEE Inter- national Conference on Distributed Computing Systems, ICDCS 2021, Washington DC, USA, July 7-10, 2021. IEEE, 875–886. doi:10.1109/ICDCS51616.2021.00088

    arXiv 2021

  3. [3]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2016. Neural Machine Translation by Jointly Learning to Align and Translate. arXiv:1409.0473 [cs.CL]

    arXiv 2016

  4. [4]

    Kunlin Cai, Jinghuai Zhang, Zhiqing Hong, William Shand, Guang Wang, Desh- eng Zhang, Jianfeng Chi, and Yuan Tian. 2024. Where Have You Been? A Study of Privacy Risk for Point-of-Interest Recommendation. InProceedings of the 30th ACM Conference on Knowledge Discovery and Data Mining(Barcelona, Spain) (KDD ’24). ACM, New York, NY, USA, 175–186. doi:10.1145...

    arXiv 2024

  5. [5]

    Carlini, C

    N. Carlini, C. Liu, Ú. Erlingsson, J. Kos, and D. Song. 2019. The Secret Sharer: Evaluating and Testing Unintended Memorization in Neural Networks. In28th USENIX Security Symposium (USENIX Security 19). 267–284

    2019

  6. [6]

    Carlini, F

    N. Carlini, F. Tramèr, E. Wallace, M. Jagielski, A. Herbert-Voss, K. Lee, A. Roberts, T. Brown, D. Song, Ú. Erlingsson, A. Oprea, and C. Raffel. 2021. Extracting Training Data from Large Language Models. In30th USENIX Security Symposium (USENIX Security 21). 2633–2650

    2021

  7. [7]

    Gobezie Chekol and M

    A. Gobezie Chekol and M. Sintayehu Fufa. 2022. A survey on next location prediction techniques, applications, and challenges.EURASIP Journal on Wireless Communications and Networking2022, 1 (2022), 29. doi:10.1186/s13638-022- 02114-6

    work page doi:10.1186/s13638-022- 2022

  8. [8]

    dataplor. 2025. Unlocking the Power of Mobility Data: Insights for Localized Strategies. https://www.dataplor.com/resources/blog/mobility-location-data/. Blog post,dataplor, 19 May 2025

    2025

Show all 44 references
  1. [9]

    Hidalgo, Michel Verleysen, and Vincent D

    Yves-Alexandre de Montjoye, César A. Hidalgo, Michel Verleysen, and Vincent D. Blondel. 2013. Unique in the Crowd: The Privacy Bounds of Human Mobility. Scientific Reports3, 1 (2013), 1376. doi:10.1038/srep01376

    2013 doi

  2. [10]

    Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng Jin. 2018. DeepMove: Predicting Human Mobility with Attentional Recurrent Networks. InProceedings of the 2018 World Wide Web Conference(Lyon, France) (WWW ’18). 1459–1468. doi:10.1145/3178876.3186058

    2018

  3. [11]

    Hana Gebrie, Hasan Farooq, and Ali Imran. 2019. What Machine Learning Predictor Performs Best for Mobility Prediction in Cellular Networks?. In2019 IEEE International Conference on Communications Workshops (ICC Workshops). 1–6. doi:10.1109/ICCW.2019.8756972

    2019

  4. [12]

    M. C. González, C. A. Hidalgo, and A. Barabási. 2008. Understanding individual human mobility patterns.Nature(2008). doi:10.1038/nature06958

    2008 doi

  5. [13]

    Xiangming Gu, Chao Du, Tianyu Pang, Chongxuan Li, Min Lin, and Ye Wang

  6. [14]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural computation9, 8 (1997), 1735–1780

    1997

  7. [15]

    Renhe Jiang, Xuan Song, Zipei Fan, Tianqi Xia, Quanjun Chen, Qi Chen, and Ryosuke Shibasaki. 2018. Deep ROI-Based Modeling for Urban Human Mobility Prediction. 2, 1, Article 14 (March 2018), 29 pages. doi:10.1145/3191746

    2018 doi

  8. [16]

    Renhe Jiang, Xuan Song, Zipei Fan, Tianqi Xia, Quanjun Chen, Satoshi Miyazawa, and Ryosuke Shibasaki. 2018. DeepUrbanMomentum: an online deep-learning system for short-term urban mobility prediction. InProceedings of the Thirty- Second AAAI Conference on Artificial Intelligenc...

    2018

  9. [17]

    Nicky Kriplani, Minh Pham, Gowthami Somepalli, Chinmay Hegde, and Niv Cohen. 2025. SolidMark: Evaluating Image Memorization in Generative Models. arXiv:2503.00592 [cs.LG]

    2025 arXiv

  10. [18]

    Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, and Nicholas Carlini. 2022. Deduplicating Training Data Makes Language Models Better. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Vo...

    2022 doi

  11. [19]

    Varshney, Mohit Bansal, Sanmi Koyejo, and Yang Liu

    Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, Kush R. Varshney, Mohit Bansal, Sanmi Koyejo, and Yang Liu. 2025. Rethinking machine unlearning for large language models.Nature Machine Int...

    2025 doi

  12. [20]

    Massimiliano Luca, Gianni Barlacchi, Bruno Lepri, and Luca Pappalardo. 2021. A Survey on Deep Learning for Human Mobility.ACM Comput. Surv.55, 1, Article 7 (Nov. 2021), 44 pages. doi:10.1145/3485125

    2021 doi

  13. [21]

    E. M. R. Oliveira, A. C. Viana, C. Sarraute, J. Brea, and I. Alvarez-Hamelin. 2016. On the regularity of human mobility.Pervasive and Mobile Computing(2016). doi:10.1016/j.pmcj.2016.04.005

    2016 doi

  14. [22]

    Mexwell. 2020. Telecom Shanghai Dataset. https://www.kaggle.com/datasets/ mexwell/telecom-shanghai-dataset. Accessed: 2025-11-13

    2020

  15. [23]

    2023.Exploring new mobil- ity policy-making pathways: Contrasting predict-and-provide and decide- and-provide

    Ministerie van Infrastructuur en Waterstaat. 2023.Exploring new mobil- ity policy-making pathways: Contrasting predict-and-provide and decide- and-provide. Technical Report TNO Publiek R12676. Dutch Ministry of Infrastructure and Water Management. https://www.government. nl/bi...

    2023

  16. [24]

    1998.Markov chains

    James R Norris. 1998.Markov chains. Number 2. Cambridge university press

    1998

  17. [25]

    Apostolos Pyrgelis, Carmela Troncoso, and Emiliano De Cristofaro. 2018. Knock Knock, Who’s There? Membership Inference on Aggregate Location Data. In 25th Annual Network and Distributed System Security Symposium, NDSS 2018, San Diego, California, USA, February 18-21, 2018. The...

    2018

  18. [26]

    Shaojie Qiao, Dayong Shen, Xiaoteng Wang, Nan Han, and William Zhu. 2015. A Self-Adaptive Parameter Selection Trajectory Prediction Approach via Hidden Markov Models.Trans. Intell. Transport. Syst.16, 1 (Jan. 2015), 284–296. doi:10. 1109/TITS.2014.2331758

    2015

  19. [27]

    Yuanyuan Qiao, Zhongwei Si, Yanting Zhang, Fehmi Ben Abdesslem, Xinyu Zhang, and Jie Yang. 2018. A hybrid Markov-based model for human mobility prediction.Neurocomputing278 (2018), 99–109. Recent Advances in Machine Learning for Non-Gaussian Data Processing. doi:10.1016/j.neuc...

    2018 doi

  20. [28]

    Xuan Rao, Lisi Chen, Yong Liu, Shuo Shang, Bin Yao, and Peng Han. 2022. Graph- Flashback Network for Next Location Recommendation. InProceedings of the 28th Auditing Memorization in Mobility Prediction Models Algorithm 1Cluster Enrichment for Reference Set Balancing Require: T...

    2022

  21. [29]

    Ragil Saputra, Suprapto, and Agus Sihabuddin. 2024. Mobility Prediction Using Markov Models: A Survey. In2024 7th International Conference on Informatics and Computational Sciences (ICICoS). 508–513. doi:10.1109/ICICoS62600.2024. 10636860

    2024

  22. [30]

    Kotz, Ravi Jain, and Xiaoning He

    Libo Song, D. Kotz, Ravi Jain, and Xiaoning He. 2006. Evaluating Next-Cell Predictors with Extensive Wi-Fi Mobility Data.IEEE Transactions on Mobile Computing5, 12 (2006), 1633–1649. doi:10.1109/TMC.2006.185

    2006 doi

  23. [31]

    Ke Sun, Tieyun Qian, Tong Chen, Yile Liang, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2020. Where to Go Next: Modeling Long- and Short-Term User Preferences for Point-of-Interest Recommendation.Proceedings of the AAAI Conference on Artificial Intelligence34, 01 (Apr. 2020), 214–...

    2020 doi

  24. [32]

    Teixeira, J

    D. Teixeira, J. Almeida, and A. C. Viana. 2021. On estimating the predictability of human mobility: the role of routine.EPJ Data Science(2021). doi:10.1140/epjds/ s13688-021-00304-8

    2021 doi

  25. [33]

    TensorFlow Authors. 2021. TensorFlow Privacy. https://github.com/tensorflow/ privacy. Includes exposure metric implementation for Secret Sharer privacy tests

    2021

  26. [34]

    Uber Technologies. 2018. H3: A Hexagonal Hierarchical Geospatial Indexing System. https://h3geo.org. Accessed: 2026-04-27

    2018

  27. [35]

    Xinglei Wang, Meng Fang, Zichao Zeng, and Tao Cheng. 2024. Where Would I Go Next? Large Language Models as Human Mobility Predictors. arXiv:2308.15197 [cs.AI]

    2024 arXiv

  28. [36]

    Jiaheng Wei, Yanjun Zhang, Leo Yu Zhang, Ming Ding, Chao Chen, Kok-Leong Ong, Jun Zhang, and Yang Xiang. 2025. Memorization in Deep Learning: A Survey. ACM Comput. Surv.58, 4, Article 98 (Oct. 2025), 35 pages. doi:10.1145/3769076

    2025 doi

  29. [37]

    2023.YJMob100K: City-Scale and Longitudinal Dataset of Anonymized Human Mobility Trajectories

    Takahiro Yabe, Kota Tsubouchi, Toru Shimizu, Yoshihide Sekimoto, Kaoru Sezaki, Esteban Moro, and Alex Pentland. 2023.YJMob100K: City-Scale and Longitudinal Dataset of Anonymized Human Mobility Trajectories. doi:10.5281/zenodo.10142719

    2023 doi

  30. [38]

    Yonga, Anne J

    Gaelle M. Yonga, Anne J. Kouam, Aline C. Viana, and Auguste V. Noumsi. 2025. On Assessing Usability and Reliability of Anonymized Spatio-Temporal Data. In2025 21st International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT). 689–69...

    2025

  31. [39]

    Desheng Zhang, Juanjuan Zhao, Fan Zhang, and Tian He. 2014. Description for Urban Data Release V2. https://people.cs.rutgers.edu/~dz220/data.html. Accessed: 2025-03

    2014

  32. [40]

    Desheng Zhang, Juanjuan Zhao, Fan Zhang, and Tian He. 2015. UrbanCPS: a cyber-physical system based on multi-source big infrastructure data for het- erogeneous model integration. InProceedings of the ACM/IEEE Sixth Interna- tional Conference on Cyber-Physical Systems. ACM, Sea...

    2015

  33. [41]

    Bob Zheng, Dhruv Ghulati, Manoj Panikkar, and Michael (Yichuan) Cai. 2025. Forecasting Models to Improve Driver Availability at Airports. https://www.uber. com/en-FR/blog/forecasting-models-to-improve-availability-at-airports/. Blog post,Uber Engineering, 19 Aug 2025

    2025

  34. [42]

    Yu Zheng. 2015. Trajectory Data Mining: An Overview.ACM Trans. Intell. Syst. Technol.6, 3, Article 29 (May 2015), 41 pages. doi:10.1145/2743025

    2015 doi

  35. [43]

    Xiaofeng Zhong, Yinfeng Xiang, Fang Yi, Chao Li, and Qinmin Yang. 2024. HMP- LLM: Human Mobility Prediction Based on Pre-trained Large Language Models. In2024 IEEE 4th International Conference on Digital Twins and Parallel Intelligence (DTPI). 687–692. doi:10.1109/DTPI61353.20...

    2024

  36. [2025]

    arXiv:2310.02664 [cs.LG]

    On Memorization in Diffusion Models. arXiv:2310.02664 [cs.LG]

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.