Pith. sign in

REVIEW 3 major objections 5 minor 57 references

This paper shows that encoding sparse mobile usage traces as correlation images lets a diffusion model generate synthetic traces nearly indistinguishable from real ones.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 02:38 UTC pith:TGSMHAAM

load-bearing objection A plausible application paper whose central metric hinges on an unvalidated heuristic decode; worth refereeing, but the inverse C-GASF step needs real validation before the strong numbers can be trusted. the 3 major comments →

arxiv 2607.14249 v1 pith:TGSMHAAM submitted 2026-07-15 cs.LG

MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion

Yilai Liu , Shiyuan Zhang , Hongyang Du This is my paper
classification cs.LG
keywords mobile usage trace generationdiffusion modelsparse multivariate time seriesCross-Gramian Angular Sum Fieldlong-tail app usagesynthetic data augmentationcorrelation imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tackles three obstacles to generating realistic mobile usage data: activity is sparse and intermittent, variables are heterogeneous (discrete app labels, continuous traffic volume, discrete location clusters), and app usage follows a long-tail distribution. It proposes C-GASF, an imaging transform that maps each timestep's app, traffic volume, and location cluster into a fixed block pattern in a correlation image, decoupling heterogeneous variables and making sparse events visible to convolutional networks. A diffusion model with attention factorized along temporal, variate, and channel axes learns this image distribution, and an inverse transform decodes generated images back into traces. On a one-week real dataset, the paper reports the best fidelity scores, with a discriminative accuracy of 0.1526 versus 0.3476 for the strongest baseline, and shows that augmenting real data with synthetic traces improves downstream traffic and location prediction. The payoff is that privacy-restricted mobile traces could be replaced by abundant, realistic synthetic ones.

Core claim

The central claim is that a two-stage imaging-diffusion pipeline can generate user-level mobile usage traces that preserve sparsity, cross-variate correlations, and long-tail app frequencies. In the encoding stage, C-GASF converts a trace into a T x C(C'+1) image: each row is one timestep, app categories and location clusters occupy separate axes, the active app-location pair creates an interest point whose value is the normalized traffic volume, inactive timesteps map to a dedicated cluster, and all other entries are fixed constants. In the generation stage, a diffusion model with factorized attention learns this image distribution. In the decoding stage, an inverse transform locates the ac

What carries the argument

The central object is the Cross-Gramian Angular Sum Field (C-GASF), a phase-correlation imaging transform that encodes one timestep of mobile usage as a C x (C'+1) matrix. The transform guarantees at most one active app and one active location cluster per timestep, placing the normalized traffic volume at their unique intersection, fixing all other entries to constants, and assigning inactive timesteps a dedicated column. This turns sparse, heterogeneous traces into a 2D image with a clean block structure that convolutional layers can learn. The generation side is a diffusion model based on a U-Net whose attention is factorized along the height (temporal), width (app/location), and channel a

Load-bearing premise

The load-bearing premise is that the inverse C-GASF decoding recovers the true active app, location cluster, and traffic volume from generated images: the decoder uses heuristic choices of the largest and most frequent image values, and the paper reports no reconstruction-error analysis on decoded outputs, so any systematic deviation in generated image patterns would propagate into all fidelity and downstream metrics.

What would settle it

Take real traces, encode them with C-GASF, run the inverse transform immediately, and measure exact recovery of app labels, location clusters, and traffic volumes. If recovery is not near-perfect, the decoding step is lossy, and reported generation metrics would not reflect the fidelity of the decoded traces.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Synthetic traces can serve as privacy-safe augmentation: in the paper's experiments, adding generated traces to the real training set improves traffic and location prediction across multiple backbone predictors, with explained variance roughly doubling for recurrent and feed-forward models.
  • Long-tail app categories remain represented rather than collapsed, so generated data could support recommendation and behavior modeling for rare apps.
  • Because the decoder uses a fixed one-active-app/one-active-location pattern, the generated data preserves the correlation between app categories and location clusters, which the paper shows baselines miss.
  • The state-of-the-art discriminative accuracy implies a trained classifier cannot easily tell synthetic traces from real ones, a direct corollary of the paper's fidelity claim.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • [Editorial inference] If C-GASF is robust, the same encode-diffuse-decode recipe could generalize to other sparse multivariate event logs—web browsing, sensor events, health records—where one category is active at a time with an attached continuous magnitude.
  • [Editorial inference] The decoder's reliance on heuristic choices of the largest and most frequent image values means reconstruction errors are most likely when generated images are soft or contain multiple near-active entries; a direct real→encode→decode reconstruction benchmark would expose this.
  • [Editorial inference] The encoding's guarantee of exactly one active app per timestep may prevent straightforward extension to overlapping or simultaneous app usage, which is common in real multitasking.
  • [Editorial inference] The paper's downstream augmentation experiment is the more practical test of value; a broader evaluation across more downstream tasks and privacy metrics would clarify how much of the fidelity advantage translates to deployed utility.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces MIDiff, a diffusion-based generative model for user-level mobile usage traces. It encodes each trace (app category, traffic volume, location/POI cluster) into a single C-GASF image via a hand-designed block-correlation pattern (Sec. 4.1), trains an unconditional U-Net diffusion model with Triplet Attention to generate such images (Sec. 4.2), and decodes generated images back into traces with a heuristic inverse transform (Appendix A). The experiments on a one-week App Usage Dataset compare against nine baselines and report SOTA DA (0.1526 vs 0.3476 for the strongest baseline), best VDS and FDDS, competitive predictive score, plus feature-level and downstream-augmentation results. The paper claims that the three central challenges—temporal sparsity, cross-channel heterogeneity, and long-tail app usage imbalance—are jointly addressed.

Significance. Assuming the reported results can be substantiated, this is a meaningful practical contribution: the C-GASF representation is an interesting way to unify discrete and continuous variates in an image space while preserving trace-level structure, and the Triplet Attention design is sensible. The authors provide code, use a real dataset, and include a broad baseline comparison; the DA improvement is large. The main unresolved point is the validation of the inverse transform. All empirical claims are computed after this heuristic decode, so the current evidence is conditional. This is fixable within the scope of the paper, but it is the central gate that determines whether the reported quantitative gains are actually attributable to the generative model.

major comments (3)
  1. [Appendix A (Algorithm 1); Sec. 5.2.2] All fidelity and utility metrics are computed after decoding generated C-GASF images using Algorithm 1. The decoder assumes that generated images exactly match the block structure of Eq. (15): one active app row with value v at the mode column, one active location column, and all other entries equal to 0 or -sqrt(1-v^2). The implemented decode is heuristic (argmax over columns per row, Mode across rows, row-wise averages after excluding columns) and will silently select the wrong app, location, or traffic if the denoised image deviates even mildly from this pattern. The paper reports no encode-decode reconstruction rate on real held-out traces, no validity statistics for generated images (e.g., what fraction of generated images satisfy the assumed block pattern), and no comparison with a soft or projection-free decoder. Consequently, the strong DA/VDS/FDDS numbers could reflect the decod
  2. [Table 3; Sec. 5.2.1] VDS and FDDS, the metrics on which the paper claims state-of-the-art performance, are reported as point values with no error bars. The text says DA is reported as mean and standard deviation over 5 iterations, but Table 3 lists only DA scores without uncertainty, and Table 4 has no uncertainty either. With FDDS, the closest baseline (Diffusion-TS) differs by only 0.0059 (0.0352 vs 0.0293); without standard errors or a confidence interval, it is unclear whether this gap is meaningful. Please report mean±std for all metrics or provide bootstrapped intervals.
  3. [Sec. 5.5 (Table 7)] The ablation claims that only diffusion with Triplet Attention can generate valid C-GASF patterns. The table reports 'GAN All Variants Failed to converge' and 'VAE [basic] Failed to form C-GASF patterns' with no training details, capacity choices, or convergence criteria. This is not sufficient to conclude that diffusion is necessary; the failure could be an artifact of the specific implementations rather than a fundamental limitation. Please provide implementation details and, ideally, a brief description of the observed failure modes.
minor comments (5)
  1. [Abstract and Sec. 4.2] The module is called 'Triplet Attention' in Sec. 4.2 but 'Triple Attention' in the abstract and Figure 4; please unify terminology.
  2. [Table 7] The table layout is garbled, with cell contents running together (e.g., 'All VariantsFailed', '✓-Failed to form C-GASF patterns✓ ✓'). Please reformat for readability.
  3. [Eq. (25)] The notation '0 = clf(\hat{x}_n)' and '1 = clf(x_n)' is unusual; please write the DA formula in the standard form |0.5 - accuracy| to avoid ambiguity.
  4. [Sec. 5.2.1] VDS and FDDS are defined per dimension or pair, but the manuscript does not state how the JSD is computed on discrete/count-valued distributions (e.g., binning scheme). Please clarify.
  5. [References] Reference [48] is a GitHub repository; if this is not the canonical publication for the ZITS method, please cite the peer-reviewed source as well.

Circularity Check

0 steps flagged

No circularity: the C-GASF/MIDiff derivation is self-contained, and the disclosed self-citation is not load-bearing.

full rationale

The paper's derivation chain is explicit and non-circular: raw traces are mapped to C-GASF images by deterministic equations (Eqs. 7-16), a diffusion model learns to denoise those images (Eqs. 17-24), and Algorithm 1 inverts the image back to traces. No equation is defined in terms of a quantity it is later claimed to predict, and no fitted parameter is renamed as a prediction. The headline metrics (DA, VDS, FDDS, predictive score) are external comparisons between generated and real traces, not quantities forced by construction. The only self-citation, [30], appears in the sentence 'This paper extends our earlier work [30]' and is descriptive; it does not supply a uniqueness theorem, an ansatz, or the central empirical claim. External citations are used for standard components (U-Net, Triplet Attention, diffusion schedules, GASF). The inverse C-GASF decode does rely on the generated image matching the structured block pattern and on using per-app maxima from the dataset, but this is a reconstruction-fidelity/validation risk rather than a circular reduction: Algorithm 1 is not a fitted estimator of the evaluation metrics, and the paper does not define C-GASF in terms of its own generated outputs. Under the requirement to exhibit a specific equation-level reduction, none exists, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central contribution is an encoding plus a learned generator on that encoding. The free parameters are small in number, but the strong invertibility assumption of C-GASF is the biggest uncharged debt: the paper assumes real traces fit the one-app-per-step pattern and that denoised images decode exactly, without error analysis.

free parameters (2)
  • C' (number of POI clusters) = 6
    Section 5.1: 'we apply k-means ... six discrete cluster labels, plus an additional label to denote inactivity timesteps.' This cluster count is chosen by the authors and directly defines the width W = C(C'+1) of the C-GASF image; results depend on this discretization but no sensitivity analysis is given.
  • T (window length) = 192
    Two-day windows at 15-minute granularity; chosen by hand and shapes the H dimension. Less critical than C', but still an experimental choice.
axioms (5)
  • domain assumption At each active timestep exactly one app category is active and its traffic volume is the only non-zero entry in the app-traffic matrix; otherwise the timestep is inactive.
    Section 4.1.1: 'constraint of having at most one non-zero entry in A~ at each active timestep is crucial, as it makes the inverse transformation possible'. The paper does not verify that real 15-minute aggregated records always satisfy this; if a user launches multiple apps in an interval, the encoding collapses them.
  • domain assumption K-means clustering of POI count vectors into C'=6 clusters yields semantically meaningful spatial contexts that are stable for generation.
    Section 5.1; no cluster-quality or stability analysis is reported.
  • domain assumption Gaussian diffusion on the C-GASF image space can learn the distribution of a structured image whose entries have special values (-1, 0, positive volumes) and whose inverse transform is valid on denoised outputs.
    Section 4.2 and Appendix A; the forward noising treats the image as continuous and does not enforce the C-GASF constraints, so the inverse transform's correctness relies on the diffusion output being near the exact pattern.
  • standard math Standard DDPM forward/reverse process and L_simple objective (Eqs. 17-24) are valid for this image distribution.
    Standard diffusion result, invoked via cosine schedule [46].
  • standard math cos(arccos x + arccos y) = xy - sqrt(1-x^2)sqrt(1-y^2).
    Trigonometric identity used in Eq. (12)-(13) to derive the C-GASF correlation matrix.

pith-pipeline@v1.3.0-alltime-deepseek · 18471 in / 14487 out tokens · 154951 ms · 2026-08-02T02:38:54.687006+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion." pith.science (2026). https://pith.science/paper/TGSMHAAM

@misc{pith2026260714249,
  author       = {Pith},
  title        = {Pith review of: MIDiff: Tackling Sparsity and Imbalance in Mobile Usage Generation via Multivariate-Imaging Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TGSMHAAM}},
  note         = {Machine review of arXiv:2607.14249}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Mobile usage traces are critical for tasks such as user behavior prediction and app recommendation, yet their use is constrained by privacy restrictions and costly large-scale data collection. Although generative models perform well on general time series, their application to mobile usage data remains challenging because (i) limited user activity causes severe sparsity, (ii) heterogeneous variable types complicate joint modeling, and (iii) functional differences across apps create pronounced usage imbalance. To address these challenges, we propose Multivariate-Imaging Diffusion (MIDiff), a diffusion-based framework operating in an imaging space defined by Cross-Gramian Angular Sum Field (C-GASF). C-GASF transforms sparse multivariate sequences into correlation images, while MIDiff employs Triple Attention in a U-Net to preserve temporal consistency and variable dependencies. Experiments show that MIDiff achieves state-of-the-art performance across fidelity metrics. In particular, it obtains a Discriminative Accuracy (DA) of 0.1526, compared with 0.3476 for the strongest baseline, ZITS-VAE, demonstrating its effectiveness in generating realistic and diverse mobile usage traces. Our code is available at https://github.com/YilaiLiu-HKU/MIDiff.

Figures

Figures reproduced from arXiv: 2607.14249 by Hongyang Du, Shiyuan Zhang, Yilai Liu.

Figure 1
Figure 1. Figure 1: Three main challenges in mobile user trace generation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The framework of MIDiff. GASF is defined by computing the trigonometric sum be￾tween every pair of angular values GASFi,j = cos(ϕi + ϕj ). (3) In the matrix, the main diagonal preserves the self￾angular information of each timestamp: GASFi,i = cos(2ϕi) = 2˜s 2 i − 1. (4) Therefore, under x˜i ∈ [0, 1], the normalized value can be recovered from the diagonal by s˜i = r GASFi,i + 1 2 . (5) The original value … view at source ↗
Figure 3
Figure 3. Figure 3: The illustration of C-GASF image. Each row contains the information of the user’s mobile usage trace in one timestep. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: MIDiff model architecture. hubs, commercial centers, at each location. To enable effec￾tive modeling and joint representation with app-traffic data, we first apply K-means clustering to these location vectors, grouping them into C ′ distinct spatial-context patterns. And subsequently we encode cluster assignments into a one-hot matrix P′(u) ∈ {0, 1} T ×(C ′+1) , P ′(u) t,j = ( 1,  j = 0 ∧ A˜ (u) t,: = 0 … view at source ↗
Figure 5
Figure 5. Figure 5: Manifold comparison of generated datasets: (a) t-SNE visualization and (b) UMAP visualization. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The heatmap illustrates different used apps in clusters of locations in different datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A comparison of the app usage distribution between [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 19 linked inside Pith

  1. [1]

    User-oriented virtual mobile network resource management for vehicle com- munications,

    H. Lu, Y. Zhang, Y. Li, C. Jiang, and H. Abbas, “User-oriented virtual mobile network resource management for vehicle com- munications,”IEEE transactions on intelligent transportation systems, vol. 22, no. 6, 2020

    2020

  2. [2]

    No-pain no-gain: Drl assisted optimization in energy-constrained cr-noma networks,

    Z. Ding, R. Schober, and H. V . Poor, “No-pain no-gain: Drl assisted optimization in energy-constrained cr-noma networks,” IEEE Transactions on Communications, vol. 69, no. 9, 2021

    2021

  3. [3]

    Understanding and prediction of mobile application usage for smart phones,

    C. Shin, J.-H. Hong, and A. K. Dey, “Understanding and prediction of mobile application usage for smart phones,” inProceedings of the 2012 ACM Conference on Ubiquitous Computing, ser. UbiComp ’12. Association for Computing Machinery, 2012

    2012

  4. [4]

    Deep learning on network traffic prediction: Recent advances, analysis, and future directions,

    O. Aouedi, V . A. Le, K. Piamrat, and Y. Ji, “Deep learning on network traffic prediction: Recent advances, analysis, and future directions,”ACM Comput. Surv., vol. 57, no. 6, Feb. 2025

    2025

  5. [5]

    Practical gan-based synthetic ip header trace generation using netshare,

    Y. Yin, Z. Lin, M. Jin, G. Fanti, and V . Sekar, “Practical gan-based synthetic ip header trace generation using netshare,” inProceedings of the ACM SIGCOMM 2022 Conference, 2022, pp. 458–472

    2022

  6. [6]

    pcapstego: A tool for generating traffic traces for experimenting with network covert channels,

    M. Zuppelli and L. Caviglione, “pcapstego: A tool for generating traffic traces for experimenting with network covert channels,” inProceedings of the 16th International Conference on Availability, Reliability and Security, 2021, pp. 1–8

    2021

  7. [7]

    Generative, high-fidelity network traces,

    X. Jiang, S. Liu, A. Gember-Jacobson, P . Schmitt, F. Bronzino, and N. Feamster, “Generative, high-fidelity network traces,” in Proceedings of the 22nd ACM Workshop on Hot Topics in Networks, 2023, pp. 131–138

    2023

  8. [8]

    Data driven prediction models of energy use of appliances in a low-energy house,

    L. M. Candanedo, V . Feldheim, and D. Deramaix, “Data driven prediction models of energy use of appliances in a low-energy house,”Energy and Buildings, vol. 140, 2017

    2017

  9. [9]

    Time-series generative adversarial networks,

    J. Yoon, D. Jarrett, and M. van der Schaar, “Time-series generative adversarial networks,” inProceedings of the 33rd International Con- ference on Neural Information Processing Systems. Curran Associates Inc., 2019

    2019

  10. [10]

    TTS-GAN: A transformer-based time-series generative adversarial network,

    X. Li, V . Metsis, H. Wang, and A. H. H. Ngu, “TTS-GAN: A transformer-based time-series generative adversarial network,” CoRR, vol. abs/2202.02691, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Mobile phone use as sequential processes: From discrete behaviors to sessions of behaviors and trajectories of sessions,

    T.-Q. Peng and J. J. H. Zhu, “Mobile phone use as sequential processes: From discrete behaviors to sessions of behaviors and trajectories of sessions,”Journal of Computer-Mediated Communica- tion, vol. 25, no. 2, pp. 129–146, Mar. 2020

    2020

  12. [12]

    Circadian pattern and burstiness in mobile phone communication,

    H.-H. Jo, M. Karsai, J. Kert ´esz, and K. Kaski, “Circadian pattern and burstiness in mobile phone communication,”New Journal of Physics, vol. 14, no. 1, p. 013055, Jan. 2012

    2012

  13. [13]

    Zero-inflated poisson regression, with an application to defects in manufacturing,

    D. Lambert, “Zero-inflated poisson regression, with an application to defects in manufacturing,”Technometrics, vol. 34, no. 1, pp. 1–14, 1992

    1992

  14. [14]

    Timevae: A variational auto-encoder for multivariate time series generation,

    A. Desai, C. Freeman, Z. Wang, and I. Beaver, “Timevae: A variational auto-encoder for multivariate time series generation,”

  15. [15]

    Diffusion-TS: Interpretable diffusion for general time series generation,

    X. Yuan and Y. Qiao, “Diffusion-TS: Interpretable diffusion for general time series generation,” inThe Twelfth International Con- ference on Learning Representations, 2024

    2024

  16. [16]

    Knowledge enhanced gan for iot traffic generation,

    S. Hui, H. Wang, Z. Wang, X. Yang, Z. Liu, D. Jin, and Y. Li, “Knowledge enhanced gan for iot traffic generation,” inProceed- ings of the ACM Web Conference 2022, ser. WWW ’22, 2022

    2022

  17. [17]

    Spatio-temporal knowledge driven diffusion model for mobile traffic generation,

    H. Chai, X. Qi, and Y. Li, “Spatio-temporal knowledge driven diffusion model for mobile traffic generation,”IEEE Transactions on Mobile Computing, vol. 24, no. 6, 2025. 12

    2025

  18. [18]

    Netdiff: A service-guided hierarchical diffusion model for network flow trace generation,

    S. Zhang, T. Li, D. Jin, and Y. Li, “Netdiff: A service-guided hierarchical diffusion model for network flow trace generation,” Proc. ACM Netw., vol. 2, no. CoNEXT3, Aug. 2024

    2024

  19. [19]

    Appgen: Mobility-aware app usage behavior generation for mobile users,

    Z. Huang, T. Li, and Y. Li, “Appgen: Mobility-aware app usage behavior generation for mobile users,”ArXiv, vol. abs/2412.07267, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Multi-time attention networks for irregularly sampled time series,

    S. N. Shukla and B. M. Marlin, “Multi-time attention networks for irregularly sampled time series,”arXiv preprint arXiv:2101.10318, 2021

    Pith/arXiv arXiv 2021

  21. [21]

    Self-supervised transformer for sparse and irregularly sampled multivariate clinical time-series,

    S. Tipirneni and C. K. Reddy, “Self-supervised transformer for sparse and irregularly sampled multivariate clinical time-series,” ACM Transactions on Knowledge Discovery from Data (TKDD), vol. 16, no. 6, pp. 1–17, 2022

    2022

  22. [22]

    Primenet: Pre-training for irregular multivariate time series,

    R. R. Chowdhury, J. Li, X. Zhang, D. Hong, R. K. Gupta, and J. Shang, “Primenet: Pre-training for irregular multivariate time series,” inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 37, no. 6, 2023, pp. 7184–7192

    2023

  23. [23]

    Addressing imbalanced domain-incremental learning through dual-balance collaborative experts,

    L. Li, D.-W. Zhou, H.-J. Ye, and D.-C. Zhan, “Addressing imbalanced domain-incremental learning through dual-balance collaborative experts,” 2025. [Online]. Available: https://arxiv. org/abs/2507.07100

    Pith/arXiv arXiv 2025

  24. [24]

    Mtlcomb: multi- task learning combining regression and classification tasks for joint feature selection,

    H. Cao, S. Rajan, B. Hahn, E. Kocak, D. Durstewitz, E. Schwarz, and V . Schneider-Lindner, “Mtlcomb: multi- task learning combining regression and classification tasks for joint feature selection,” 2024. [Online]. Available: https: //arxiv.org/abs/2405.09886

    Pith/arXiv arXiv 2024

  25. [25]

    Addressing spatial-temporal heterogeneity: General mixed time series analysis via latent continuity recovery and align- ment,

    J. Chen, “Addressing spatial-temporal heterogeneity: General mixed time series analysis via latent continuity recovery and align- ment,”Advances in Neural Information Processing Systems, vol. 37, pp. 17 910–17 946, 2024

    2024

  26. [26]

    General-purpose user embeddings based on mobile app usage,

    J. Zhang, B. Bai, Y. Lin, J. Liang, K. Bai, and F. Wang, “General-purpose user embeddings based on mobile app usage,” inProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ser. KDD ’20. New York, NY, USA: Association for Computing Machinery, 2020, pp. 2831–2840. [Online]. Available: https://doi.org/10.1145/339...

    arXiv 2020

  27. [27]

    Real world longitudinal ios app usage study at scale,

    D. Kim, J. Gluck, M. Hall, and Y. Agarwal, “Real world longitudinal ios app usage study at scale,” 2019. [Online]. Available: https://arxiv.org/abs/1912.12526

    Pith/arXiv arXiv 2019

  28. [28]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical Im- age Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer International Publishing, 2015, pp. 234–241

    2015

  29. [29]

    Rotate to attend: Convolutional triplet attention module,

    D. Misra, T. Nalamada, A. U. Arasanipalai, and Q. Hou, “Rotate to attend: Convolutional triplet attention module,”2021 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020

    2021

  30. [30]

    Y. Liu, S. Zhang, and H. Du,Poster: Enhancing Mobile Traffic Data Generation through Spatio-temporal Correlation Imaging. Association for Computing Machinery, 2025

    2025

  31. [31]

    Smartphone app usage prediction using points of interest,

    D. Yu, Y. Li, F. Xu, P . Zhang, and V . Kostakos, “Smartphone app usage prediction using points of interest,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 1, no. 4, 2018

    2018

  32. [32]

    Generating multivariate time series with common source coordinated gan (cosci-gan),

    A. Seyfi, J.-F. Rajotte, and R. T. Ng, “Generating multivariate time series with common source coordinated gan (cosci-gan),” in Proceedings of the 36th International Conference on Neural Information Processing Systems, ser. NIPS ’22. Curran Associates Inc., 2022

    2022

  33. [33]

    Variational recurrent auto-encoders,

    O. Fabius, J. R. van Amersfoort, and D. P . Kingma, “Variational recurrent auto-encoders,”CoRR, vol. abs/1412.6581, 2014

    Pith/arXiv arXiv 2014

  34. [34]

    Causal recurrent variational autoencoder for medical time series generation,

    H. Li, S. Yu, and J. C. Pr ´ıncipe, “Causal recurrent variational autoencoder for medical time series generation,”ArXiv, vol. abs/2301.06574, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Population aware diffusion for time series generation,

    Y. Li, M. Han, Z. Bi, I. T. Urnes, and H. Chen, “Population aware diffusion for time series generation,”ArXiv, vol. abs/2501.00910, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Imaging time-series to improve classi- fication and imputation,

    Z. Wang and T. Oates, “Imaging time-series to improve classi- fication and imputation,” inProceedings of the 24th International Conference on Artificial Intelligence, ser. IJCAI’15. AAAI Press, 2015

    2015

  37. [37]

    Netdiffus: Network traffic generation by diffusion models through time-series imaging,

    N. Sivaroopan, D. Bandara, C. Madarasingha, G. Jourjon, A. Jayasumana, and K. Thilakarathna, “Netdiffus: Network traffic generation by diffusion models through time-series imaging,”

  38. [38]

    Utilizing image transforms and diffusion models for gen- erative modeling of short and long time series,

    I. Naiman, N. Berman, I. Pemper, I. Arbiv, G. Fadlon, and O. Azen- cot, “Utilizing image transforms and diffusion models for gen- erative modeling of short and long time series,” inAdvances in Neural Information Processing Systems, A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, Eds., vol. 37. Curran Associates, Inc., 2024

    2024

  39. [39]

    Timesnet: Temporal 2d-variation modeling for general time series analysis,

    H. Wu, T. Hu, Y. Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  40. [40]

    Deep transfer learning for city-scale cellular traffic generation through urban knowledge graph,

    S. Zhang, T. Li, S. Hui, G. Li, Y. Liang, L. Yu, D. Jin, and Y. Li, “Deep transfer learning for city-scale cellular traffic generation through urban knowledge graph,” inProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ser. KDD ’23. Association for Computing Machinery, 2023

    2023

  41. [41]

    Diffusion model-based mobile traffic generation with open data for network planning and op- timization,

    H. Chai, T. Jiang, and L. Yu, “Diffusion model-based mobile traffic generation with open data for network planning and op- timization,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, ser. KDD ’24. Association for Computing Machinery, 2024

    2024

  42. [42]

    Spatio-temporal diffusion model for cellular traffic generation,

    X. Liu, X. Xu, Z. Liu, Z. Li, and K. Wu, “Spatio-temporal diffusion model for cellular traffic generation,”IEEE Transactions on Mobile Computing, vol. 25, no. 1, 2026

    2026

  43. [43]

    Mobile user traffic generation via multi-scale hierarchical gan,

    T. Li, S. Hui, S. Zhang, H. Wang, Y. Zhang, P . Hui, D. Jin, and Y. Li, “Mobile user traffic generation via multi-scale hierarchical gan,” ACM Trans. Knowl. Discov. Data, 2024

    2024

  44. [44]

    Packetdiff: A flow guided diffusion model for network packet trace generation,

    S. Zhang, H. Chai, Y. Li, B. Qiu, L. Yue, and R. Pan, “Packetdiff: A flow guided diffusion model for network packet trace generation,” IEEE Internet of Things Journal, 2025

    2025

  45. [45]

    Lsdm: Llm- enhanced spatio-temporal diffusion model for service-level mobile traffic prediction,

    S. Zhang, T. Li, Z. Xiao, H. Du, and K. Huang, “Lsdm: Llm- enhanced spatio-temporal diffusion model for service-level mobile traffic prediction,”arXiv preprint arXiv:2507.17795, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Improved denoising diffusion proba- bilistic models,

    A. Nichol and P . Dhariwal, “Improved denoising diffusion proba- bilistic models,”ArXiv, vol. abs/2102.09672, 2021

    Pith/arXiv arXiv 2021

  47. [47]

    Timeautodiff: A unified framework for generation, imputation, forecasting, and time-varying metadata conditioning of heterogeneous time series tabular data,

    N. Suh, Y. Yang, D.-Y. Hsieh, Q. Luan, S. Xu, S. Zhu, and G. Cheng, “Timeautodiff: A unified framework for generation, imputation, forecasting, and time-varying metadata conditioning of heterogeneous time series tabular data,” 2025. [Online]. Available: https://arxiv.org/abs/2406.16028

    arXiv 2025

  48. [48]

    Zero-inflated time series gen- eration,

    E.-R. Ardelean, “Zero-inflated time series gen- eration,” https://github.com/ArdeleanRichard/ Zero-Inflated-Time-Series-Generation, 2026, accessed: 2026-06-27

    2026

  49. [49]

    Tsgbench: Time series generation benchmark,

    Y. Ang, Q. Huang, Y. Bao, A. K. H. Tung, and Z. Huang, “Tsgbench: Time series generation benchmark,”ArXiv, vol. abs/2309.03755, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    Generative adversarial networks,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde- Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” 2014. [Online]. Available: https://arxiv. org/abs/1406.2661

    Pith/arXiv arXiv 2014

  51. [51]

    Auto-encoding variational bayes,

    D. P . Kingma and M. Welling, “Auto-encoding variational bayes,”

  52. [52]

    Timexer: Empowering transformers for time series forecasting with exogenous variables,

    Y. Wang, H. Wu, J. Dong, G. Qin, H. Zhang, Y. Liu, Y. Qiu, J. Wang, and M. Long, “Timexer: Empowering transformers for time series forecasting with exogenous variables,” 2024. [Online]. Available: https://arxiv.org/abs/2402.19072

    Pith/arXiv arXiv 2024

  53. [53]

    Temporal query network for efficient multivariate time series forecasting,

    S. Lin, H. Chen, H. Wu, C. Qiu, and W. Lin, “Temporal query network for efficient multivariate time series forecasting,” 2025. [Online]. Available: https://arxiv.org/abs/2505.12917

    Pith/arXiv arXiv 2025

  54. [54]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” 2024. [Online]. Available: https: //arxiv.org/abs/2312.00752 13 APPENDIXA INVERSETRANSFORM OFCROSS-GRAMIANANGU- LARSUMFIELD Algorithm 1Single Image Decoding Require:ImageI∈R 1×T×C(C ′+1), timestepsT, the num- ber of PoI clustersC ′ and app categoriesC.l max,i = maxu∈U, τ...

    Pith/arXiv arXiv 2024

  55. [2021]

    Available: https://arxiv.org/abs/2111.08095

    [Online]. Available: https://arxiv.org/abs/2111.08095

    Pith/arXiv arXiv

  56. [2022]

    Available: https://arxiv.org/abs/1312.6114

    [Online]. Available: https://arxiv.org/abs/1312.6114

    Pith/arXiv arXiv

  57. [2023]

    Available: https://arxiv.org/abs/2310.04429

    [Online]. Available: https://arxiv.org/abs/2310.04429

    Pith/arXiv arXiv