A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations

Peng Hu; Xiang Shi

REVIEW 4 major objections 5 minor 18 references

A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims that GT-GAN, an autoencoder-wrapped GAN, produces realistic LEO satellite-network data from incomplete inputs and remains the most accurate of three generators when 40% of observations are missing.

desk verdict First application-level benchmark of GenAI time-series synthesis on incomplete LEO satellite data; the robustness ranking is plausible but rests on a small single-site sample and an artificial block-missing mask. read the letter →

arxiv 2607.24790 v1 pith:XULKYJWR submitted 2026-06-28 cs.AI cs.LGcs.NI

Xiang Shi , Peng Hu This is my paper · ORCID

classification cs.AIcs.LGcs.NI

keywords generativeadversarialnetworksLEOsatelliteInternetdataaugmentationmissingtime-seriesgenerationnetworkmeasurementGT-GANrobustness

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

This paper tries to establish that a generative adversarial network can synthesize realistic LEO satellite Internet measurement data directly from incomplete observations, so data augmentation no longer requires complete records. It proposes two missing-data scenarios—consecutive block losses and random point-wise losses—and compares GT-GAN, SeriesGAN, and a temporal VAE on a two-day Starlink observation subset. The authors claim GT-GAN produces synthetic data that is hardest to distinguish from real data and most useful for downstream prediction, and that its advantage grows as missingness increases to 40%. If correct, GenAI becomes a practical augmentation layer for satellite network measurement, a domain where missing data is common.

What carries the argument

GT-GAN, a general-purpose time-series GAN named in the paper, wraps an autoencoder around the generator-discriminator pair so the generator learns global sequence-level temporal features rather than only local patterns. The paper pairs this with two empirically motivated masking mechanisms—block-wise masks that delete the first k points of each sliding window (k=5, 10, 20, 40) and point-wise masks drawn from a Bernoulli distribution (p=5%, 10%, 20%, 40%)—and measures success with discriminative score, predictive score, and t-SNE overlap. The autoencoder's global latent representation is the mechanism that lets GT-GAN survive fragmented inputs.

What would settle it

Repeat the comparison on a longer, multi-terminal LEO dataset whose missing blocks come from real recorded outages rather than synthetic masks, without manual imputation, and check whether GT-GAN still has the lowest discriminative and predictive scores at 20–40% missingness; if a rival model matches or beats it, the central robustness claim fails.

Watch

Extended reading notes

Core claim

The central claim, stated in the conclusion, is that GT-GAN is the most robust of the three tested generators for incomplete LEO satellite Internet observations: even when 40% of the input data is missing, it continues to capture the underlying data distribution and suffers the least degradation in generalization. In the block-wise scenario, GT-GAN had the lowest discriminative and predictive scores at all four missing scales. In the point-wise scenario, GT-GAN was not best at 5–10% missingness, but at 20% and 40% it was the only model whose synthetic points still overlapped the real distribution, while SeriesGAN underwent mode collapse and the temporal VAE degraded sharply.

Load-bearing premise

The ranking rests on a single two-day, 2,880-sample slice of one terminal's Starlink measurements, with two gaps manually mean-imputed, and on synthetic missing masks that may not match real-world missingness; if those masks or that slice misrepresent actual LEO data loss, the robustness ranking may not transfer.

Editorial extensions

If this is right

GT-GAN can synthesize useful training data directly from incomplete LEO satellite observations, removing the need to impute missing values before augmentation.
When 20–40% of inputs are missing, a global sequence-level GAN design is more reliable than local reconstruction-based generation, which degrades or collapses.
The proposed block-wise and point-wise missing scenarios give a reproducible benchmark for comparing generative models on satellite network data.
Downstream predictors trained on GT-GAN synthetic data achieve near-real predictive scores, suggesting the synthetic data carries usable temporal dependencies.
At low missingness (5–10%), simpler local models can match or beat GT-GAN, so the choice of generator should depend on the expected data loss rate.

Reading between the lines

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

Because the experiments use only a two-day slice from one terminal, the claimed robustness ranking is a proof of concept; the decisive test is a longer, multi-terminal dataset with naturally occurring outages.
The synthetic missing masks assume loss is either leading-block or Bernoulli point-wise; real LEO outages may be driven by weather, orbital geometry, or terminal faults, so correlated missingness is an untested regime.
The paper observes mode collapse at high missing rates and lists Wasserstein loss with gradient penalty plus diversity loss as future work; that is a concrete, testable fix for the failure it acknowledges.
A natural downstream extension is to compare GT-GAN augmentation against classical mean/interpolation imputation in a full prediction pipeline, since the TSTR metric already suggests synthetic data can stand in for real observations.

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.

Desk Editor's Note

read the letter

Here's the short version. This paper does something genuinely new at the application level—nobody else has benchmarked GT-GAN, SeriesGAN, and Temporal VAE on incomplete LEO satellite Internet observations—and the main robustness result, that GT-GAN holds up best when 40% of the input is missing, is credible for the exact experimental setups used. But the evidence base is narrow, and one of the two missing-data scenarios is artificially constructed in a way that could bias the comparison. I'd send it to peer review, but only with a clear expectation of major revision.

The strengths first. The authors use public data (WetLinks), external models, and a standard evaluation protocol (discriminative and predictive scores from TimeGAN). The point-wise missing scenario uses Bernoulli masks, which is a reasonable model for random single-point loss. The tables show a consistent pattern: at high missingness, SeriesGAN collapses (predictive scores jump to 0.25–0.36) and Temporal VAE degrades (0.15–0.32), while GT-GAN stays around 0.05. That is a real, measured difference, not a fitted claim. The t-SNE plots, though qualitative, are consistent with the numbers.

Now the soft spots, in rough order of importance. First, the block-wise missing scenario in Eq. (1) always masks the first k observations of every sliding window. That means every training window has a fully observed suffix, so a model can learn to reconstruct the prefix from later context. Real gaps in WetLinks start at arbitrary times and have varied lengths (up to 51,570 consecutive points). The stress-test note is right: this is a narrow special case, and the block-wise ranking might change if masks were placed at random positions or at the end. Second, the evaluation is one site, two days, 2,880 samples, with two gaps manually filled by mean imputation. That is a thin base for a general claim about "LEO satellite Internet observations." Third, there are no error bars or repeated runs, and no code or hyperparameters released, so the stability of the ranking is unknown. Fourth, the predictive score is defined as |1−MAE|, with lower better; that is internally inconsistent with the table values (a low |1−MAE| would mean MAE close to 1, which is bad). Likely a typo, but it needs fixing. Fifth, there is no comparison to even a simple imputation baseline, so we don't know whether the GANs beat interpolation for augmentation purposes.

None of these flaws kills the core observation. The point-wise scenario independently supports GT-GAN's robustness at high missingness, so I wouldn't call the headline an artifact. But the paper overclaims by presenting a 2-day single-site study with an unrealistic block mask as evidence about LEO networks in general. The honest conclusion is "GT-GAN looks robust for these two synthetic missing patterns; more work needed."

Who is this for? People working on ML data preparation for network measurement, especially satellite networks. It is a useful starting point, not a definitive benchmark. I'd read it, maybe present it in a group, but I wouldn't cite it as established fact. For peer review: yes, it deserves referee time, because the question is relevant and the preliminary evidence is interesting. But I'd want code, variance estimates, realistic missing masks, and a corrected metric before accepting.

Referee Report

4 major / 5 minor

Summary. The paper addresses missing data in LEO satellite Internet measurement datasets by proposing two synthetic missingness scenarios (block-wise and point-wise) and evaluating three generative models—GT-GAN, SeriesGAN, and Temporal VAE—on a two-day WetLinks subset. The authors report discriminative and predictive scores plus t-SNE visualizations, concluding that GT-GAN is the most robust and best-performing model, especially under 40% missingness. The central empirical claim is that GT-GAN degrades least as missingness intensifies, whereas the two baselines deteriorate markedly.

Significance. If the central claim holds, the paper would be a useful application-driven comparison showing that a GAN with an external autoencoder can synthesize usable LEO satellite network data from highly incomplete inputs. The work uses a public dataset (WetLinks) and externally published baseline models (SeriesGAN, Temporal VAE relative to the VAE family), and the TSTR-style evaluation gives concrete, quantitative scores. However, the significance is presently tempered by overclaims and by the narrow synthetic missingness design; the finding is not yet established as a general robustness result for LEO observations.

major comments (4)

[§IV-B, Table I] The claim that 'GT-GAN is the top-performing model of the three' and that its discriminative scores are 'significantly lower than' the others is directly contradicted by Table I: at k=10, Temporal VAE has a discriminative score of 0.3184 versus GT-GAN's 0.4339, and at k=20 Temporal VAE scores 0.4122 versus GT-GAN's 0.4825. Thus GT-GAN does not 'consistently outperform' in both metrics, and the abstract's statement that GT-GAN exhibits 'the best performance among all models in both missing scenarios' is unsupported. The authors should restrict the claim to predictive scores, or to high missingness, or provide a principled aggregation that justifies their overall ranking.
[§III-A, Eq. (1)] The block-wise missing scenario always masks the first k observations of every sliding window. This is a narrow special case: the WetLinks gaps described in §I and §IV-A have arbitrary start times and widely varying lengths, including an extreme 51,570-point gap. With prefix-only masks, every window contains a fully observed suffix, allowing models to exploit backward extrapolation from later context. This may artificially favor GT-GAN's autoencoder-based architecture. No sensitivity analysis over mask position or over the empirical gap-length/start-time distribution is provided. The block-wise robustness ranking at k=40 is therefore conditional on this artificial mask placement and does not establish the paper's general robustness conclusion.
[§IV-C, Table II] The point-wise experiment shows the predicted degradation pattern at p≥20%, but the low-missingness rows contradict the 'best among all models in both missing scenarios' claim. At p=5%, SeriesGAN has a predictive score of 0.0486 versus GT-GAN's 0.0592, and at p=10%, Temporal VAE has a discriminative score of 0.3541 versus GT-GAN's 0.4904. The paper's own narrative acknowledges that GT-GAN's advantage appears only at higher missing rates, which is compatible with the data, but the conclusion and abstract overstate the result by implying uniform superiority. The claims need to be aligned with the measured numbers.
[§IV-A and §IV-C] The robustness finding rests on a single two-day, 2,880-sample subset from one WetLinks terminal, after manual mean imputation of two gaps, with no cross-site or multi-seed evaluation. The point-wise Bernoulli scenario uses one fixed random seed, and no confidence intervals or significance tests are reported. The predictive-score differences at high missingness (e.g., 0.0502 vs. 0.3647 at p=40) are large, but the discriminative-score differences are small; without repeated runs or broader data, the paper cannot support the transferable claim that GT-GAN is 'the most robust and reliable solution' for LEO satellite Internet observations.

minor comments (5)

[§II] Reference [6] is attributed to 'Morgan et al.' but the arXiv paper is by Olof Mogren; the author name is misspelled. Also, '20219' appears to be a typo for '2019'.
[§III-A] Equation (1) uses inconsistent indexing: the mask is defined as m_i, but the condition is written with t. The range of i and the meaning of t should be clarified.
[§III-C] The predictive score is defined as |1−MAE|, but the MAE is not explicitly defined, nor is it stated whether it is computed on normalized data or in original units. This affects interpretability across tables.
[§IV-B, Figs. 1–3] The t-SNE figures have no legends; colors are described in the text but the captions do not identify which color is real data and which is synthetic. This makes the visual claims harder to verify.
[§V] The conclusion acknowledges that GT-GAN 'still experiences mode collapse under severe missing rates.' This admission is in tension with the phrasing 'consistently capturing the underlying input data distribution' in the abstract; the wording should be softened to avoid self-contradiction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical comparison of externally published models on a public dataset, with measured rather than derived metrics.

full rationale

The paper's central claim is an empirical comparison of three generative models (GT-GAN, SeriesGAN, and Temporal VAE) under two missing-data scenarios on the public WetLinks dataset. The discriminative and predictive scores in Tables I and II are measured outputs of a train/test protocol (an RNN classifier and the TSTR predictive score) adopted from TimeGAN; they are not fitted parameters and are not implied by the missing-scenario definitions. GT-GAN is an existing model from reference [9], not a construction of the current authors, so there is no self-citation chain that carries the argument. The block-wise scenario in Eq. (1) always masks the first k positions of each sliding window, which may narrow the generality of the robustness ranking, and the 2-day subset with manual imputation limits external validity, but these are evaluation-design and correctness concerns, not circularity. No prediction in the paper reduces to an input by construction, and no equation is definitionally equivalent to the result it is used to support. Therefore no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No traditional fitted physical constants appear, and the paper introduces no new entities. The load-bearing choices are experimental design parameters (window length L, missing-rate levels, VAE sizes) plus assumptions about missingness mechanisms, metric validity, and dataset representativeness. The unreported GT-GAN/SeriesGAN hyperparameters add an unlisted set of choices that could affect the ranking.

free parameters (3)

Block-wise window length L
The block-wise missing scenario divides the dataset into windows of length L, but L is never specified in the paper. The masking definition in Eq. (1) depends on L, and changing L would change how 'first k observations masked' is applied and the effective training context.
Missing-rate levels k and p = k in {5,10,20,40}; p in {5%,10%,20%,40%}
These levels are chosen by hand rather than derived from the observed missingness distribution. The central claim about robustness at '40% missing' is tied to these manually selected thresholds; other levels could change the ranking.
Temporal VAE architecture sizes = hidden 64, latent 32
The Temporal VAE baseline uses a 2-layer bidirectional GRU encoder and 2-layer GRU decoder with hidden dimension 64 and latent dimension 32. These are hand-chosen hyperparameters that affect the strength of the baseline and therefore the comparison; GT-GAN and SeriesGAN hyperparameters are not reported.

assumptions (4)

domain assumption Missingness is missing-completely-at-random and independent per time point in the point-wise scenario.
Eq. (2) draws masks from an i.i.d. Bernoulli distribution. Real LEO measurement missingness may be correlated with weather, orbital geometry, or outage type; if the missing mechanism is informative, the robustness ranking may not transfer.
domain assumption The discriminative score, predictive score, and t-SNE overlap are valid proxies for 'high-fidelity' and 'generalization' of generated data.
Section III-C adopts the TimeGAN evaluation protocol without validating these metrics against a downstream task specific to satellite network measurements. The predictive-score formula is also internally inconsistent, so the metric axiom is fragile.
domain assumption A 2-day subset (2,880 samples) from a single WetLinks Enschede terminal is representative of LEO satellite Internet observations.
Section IV-A selects October 7-8, 2023 data from one region. The conclusions in the abstract are phrased broadly about LEO satellite network data, but the evidence is from one short, single-site trace.
domain assumption GAN/VAE models can recover the underlying data distribution from incomplete training data when masks are used.
The entire framework assumes that training on artificially masked versions of a complete dataset teaches the model the complete distribution. The paper does not prove this; it only reports empirical scores.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations." pith.science (2026). https://pith.science/paper/XULKYJWR

@misc{pith2026260724790,
  author       = {Pith},
  title        = {Pith review of: A GAN-Based Framework for Robust Data Synthesis in Satellite Internet Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XULKYJWR}},
  note         = {Machine review of arXiv:2607.24790}
}

read the original abstract

Low-Earth orbit (LEO) satellite Internet has become an important infrastructure for enabling ubiquitous connectivity to align with the International Telecommunications Union vision for 6G telecommunications networks. However, current LEO satellite Internet observations often suffer from missing data, which complicates data augmentation task and limits the expansion of representative datasets. Given the complex characteristics of these datasets, generative AI (GenAI) presents a promising approach, yet its application in this domain has received little attention to date. In this paper, we propose a GenAI-based framework to synthesize high-fidelity data directly from incomplete LEO network observations. We propose the representative data missing scenarios, and evaluate the performance with the latest GAN- and VAE-based GenAI models on the recent WetLinks dataset. We design block-wise and point-wise missing scenarios to closely simulate the data loss that happens on real-world LEO satellite networks. Our results show the effectiveness of our proposed GAN-based framework and GT-GAN model exhibits the best performance among all models in both missing scenarios. Even under extreme conditions (e.g., 40% of the input data is missing), GT-GAN shows the highest robustness, consistently capturing the underlying input data distribution and being the least affected in terms of generalization. Our results shed light on future directions for GenAI-based data augmentation methods and data-driven research on satellite network measurement.

Figures

Figures reproduced from arXiv: 2607.24790 by the authors.

**Figure 2.** SeriesGAN multiple t-SNE plots for block-wise missing comparison [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Temporal VAE model multiple t-SNE plots for block-wise missing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: SeriesGAN multiple t-SNE plots for Point-wise missing scenario [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: GT-GAN model multiple t-SNE plots for Point-wise missing [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Temporal VAE model multiple t-SNE plots for point-wise missing [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 5 linked inside Pith

[1]

Lens: A leo satellite network measurement dataset,

J. Zhao and J. Pan, “Lens: A leo satellite network measurement dataset,” inProceedings of the 15th ACM Multimedia Systems Con- ference, 2024, pp. 278–284

2024
[2]

WetLinks: a Large-Scale Longitudinal Starlink Dataset with Contiguous Weather Data,

D. Laniewski, E. Lanfer, B. Meijerink, R. van Rijswijk-Deij, and N. Aschenbruck, “WetLinks: a Large-Scale Longitudinal Starlink Dataset with Contiguous Weather Data,” inProc. of the 8th Network Traffic Measurement and Analysis Conf. (TMA), 2024

2024
[3]

A multifaceted look at starlink performance,

N. Mohan, A. E. Ferguson, H. Cech, R. Bose, P. R. Renatin, M. K. Marina, and J. Ott, “A multifaceted look at starlink performance,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2723–2734

2024
[4]

Statistical characterization and prediction of e2e latency over leo satellite networks,

A. Casparsen, J. E. Jakobsen, J. J. Nielsen, P. Popovski, and I. L. Mayorga, “Statistical characterization and prediction of e2e latency over leo satellite networks,” 2026, arXiv:2601.08439 [cs.NI]

arXiv 2026
[5]

Generative adversarial nets,

I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

2014
[6]

C-rnn-gan: Continuous recurrent neural networks with adversarial training,

O. Mogren, “C-rnn-gan: Continuous recurrent neural networks with adversarial training,” 2016, arXiv:1611.09904 [cs.AI]

arXiv 2016
[7]

Conditional gan for timeseries generation,

K. E. Smith and A. O. Smith, “Conditional gan for timeseries generation,” 2020, arXiv:2006.16477 [cs.LG]

arXiv 2020
[8]

Time-series generative adversarial networks,

J. Yoon, D. Jarrett, and M. Van der Schaar, “Time-series generative adversarial networks,”Advances in neural information processing systems, vol. 32, 2019

2019

Show all 18 references

[9]

Gt-gan: General purpose time series synthesis with generative adversarial networks,

J. Jeon, J. Kim, H. Song, S. Cho, and N. Park, “Gt-gan: General purpose time series synthesis with generative adversarial networks,” Advances in Neural Information Processing Systems, vol. 35, pp. 36 999–37 010, 2022

2022
[10]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” 2013, arXiv:1312.6114 [stat.ML]

2013 arXiv
[11]

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

A. Desai, C. Freeman, Z. Wang, and I. Beaver, “Timevae: A vari- ational auto-encoder for multivariate time series generation,” 2021, arXiv:2111.08095 [cs.LG]

2021 arXiv
[12]

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

X. Yuan and Y . Qiao, “Diffusion-ts: Interpretable diffusion for general time series generation,” 2024, arXiv:2403.01742 [cs.LG]

2024 arXiv
[13]

Interpretable genai: Synthetic financial time series generation with probabilistic lstm,

C. Schwarz, “Interpretable genai: Synthetic financial time series generation with probabilistic lstm,”Available at SSRN 4877007, 2024

2024
[14]

Ts-gan: Time-series gan for sensor- based health data augmentation,

Z. Yang, Y . Li, and G. Zhou, “Ts-gan: Time-series gan for sensor- based health data augmentation,”ACM Transactions on Computing for Healthcare, vol. 4, no. 2, pp. 1–21, 2023

2023
[15]

Generation of realistic synthetic financial time-series,

M. Dogariu, L.-D. S ¸tefan, B. A. Boteanu, C. Lamba, B. Kim, and B. Ionescu, “Generation of realistic synthetic financial time-series,” ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), vol. 18, no. 4, pp. 1–27, 2022

2022
[16]

Time series data aug- mentation for energy consumption data based on improved timegan,

P. Tang, Z. Li, X. Wang, X. Liu, and P. Mou, “Time series data aug- mentation for energy consumption data based on improved timegan,” Sensors, vol. 25, no. 2, p. 493, 2025

2025
[17]

Seriesgan: Time series generation via adversarial and autoregressive learning,

M. EskandariNasab, S. M. Hamdi, and S. F. Boubrahimi, “Seriesgan: Time series generation via adversarial and autoregressive learning,” in 2024 IEEE International Conference on Big Data (BigData). IEEE, 2024, pp. 860–869

2024
[18]

Visualizing data using t-sne

L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of machine learning research, vol. 9, no. 11, 2008

2008

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Lens: A leo satellite network measurement dataset,

J. Zhao and J. Pan, “Lens: A leo satellite network measurement dataset,” inProceedings of the 15th ACM Multimedia Systems Con- ference, 2024, pp. 278–284

2024

[2] [2]

WetLinks: a Large-Scale Longitudinal Starlink Dataset with Contiguous Weather Data,

D. Laniewski, E. Lanfer, B. Meijerink, R. van Rijswijk-Deij, and N. Aschenbruck, “WetLinks: a Large-Scale Longitudinal Starlink Dataset with Contiguous Weather Data,” inProc. of the 8th Network Traffic Measurement and Analysis Conf. (TMA), 2024

2024

[3] [3]

A multifaceted look at starlink performance,

N. Mohan, A. E. Ferguson, H. Cech, R. Bose, P. R. Renatin, M. K. Marina, and J. Ott, “A multifaceted look at starlink performance,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2723–2734

2024

[4] [4]

Statistical characterization and prediction of e2e latency over leo satellite networks,

A. Casparsen, J. E. Jakobsen, J. J. Nielsen, P. Popovski, and I. L. Mayorga, “Statistical characterization and prediction of e2e latency over leo satellite networks,” 2026, arXiv:2601.08439 [cs.NI]

arXiv 2026

[5] [5]

Generative adversarial nets,

I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014

2014

[6] [6]

C-rnn-gan: Continuous recurrent neural networks with adversarial training,

O. Mogren, “C-rnn-gan: Continuous recurrent neural networks with adversarial training,” 2016, arXiv:1611.09904 [cs.AI]

arXiv 2016

[7] [7]

Conditional gan for timeseries generation,

K. E. Smith and A. O. Smith, “Conditional gan for timeseries generation,” 2020, arXiv:2006.16477 [cs.LG]

arXiv 2020

[8] [8]

Time-series generative adversarial networks,

J. Yoon, D. Jarrett, and M. Van der Schaar, “Time-series generative adversarial networks,”Advances in neural information processing systems, vol. 32, 2019

2019