Pith. sign in

REVIEW 4 major objections 5 minor 35 references

From Noise to Precision: A Diffusion-Driven Approach to Zero-Inflated Precipitation Prediction

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

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

Pith's one-line read The paper argues that zero-inflated precipitation forecasting can be improved by smoothing the target with Gaussian noise, predicting on the smoothed series, and then denoising the prediction with a diffusion model trained on clean rainfall

desk verdict ZIDF is an interesting empirical recipe, but its theoretical guarantee is unsupported and the reported gains may be shrinkage. read the letter →

arxiv 2509.10501 v1 pith:4TOU7SHM submitted 2025-09-01 cs.LG cs.AI

Wentao Gao , Jiuyong Li , Lin Liu , Thuc Duy Le , Xiongren Chen , Xiaojing Du , Jixue Liu , Yanchang Zhao
show 1 more author
Yun Chen
This is my paper
classification cs.LGcs.AI
keywords zero-inflateddataprecipitationforecastingdiffusionmodelsGaussianperturbationDDPMtimeseriesTransformerconditionalmean
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

Precipitation series are zero-inflated: most days record no rain, with occasional heavy events. The paper's proposal is to stop forcing a forecasting model to fit that discontinuous mixture directly. Instead, add Gaussian noise to the target so the point mass at zero becomes a smooth distribution, train a Transformer to predict the smoothed target, and run a diffusion model trained on clean rainfall backward from the Transformer's prediction to strip the noise away. Its formal claim (Theorem 1) is that under an unbiasedness assumption the denoised forecast equals the conditional mean of clean rainfall given history. Its empirical claim on South Australian data is a reduction of up to 56.7% in MSE and 21.1% in MAE relative to a Non-stationary Transformer baseline.

What carries the argument

The mechanism is a three-stage pipeline. Gaussian convolution replaces the Dirac delta at zero with a smooth bump, making the zero-inflated distribution infinitely differentiable. A Non-stationary Transformer is then trained by MSE on these smoothed targets; the theory requires this component to be unbiased. Third, at inference, the Transformer's noisy prediction is treated as a high-noise state of the diffusion forward process, embedded at a matching step τ, and the learned reverse diffusion computes the posterior mean E[Y|Y(τ)]. The identity doing the work is the DDPM posterior-mean property: a denoiser trained with the simple denoising objective outputs the conditional mean of the clean s

What would settle it

Run ZIDF on synthetic zero-inflated series with a known conditional mean E[Y|history], then replace the predictive model's errors with obviously non-Gaussian noise (e.g., heavy-tailed or skewed) and check whether the denoised predictions still equal the true conditional mean; if the outputs are biased toward the marginal mean, Theorem 1's key assumption is violated.

Watch

Extended reading notes

Core claim

The central discovery is a mechanism for zero-inflated time-series forecasting: rather than fitting the discontinuous mixture of a point mass at zero and a continuous rainfall distribution, one perturbs the target with Gaussian noise, forecasts the smoothed target with a Transformer, then uses a Denoising Diffusion Probabilistic Model (a neural network trained to reverse Gaussian noise added in a forward process) to recover clean targets. Under Assumption 1, which requires the predictive model to learn the conditional mean of the smoothed target, and assuming the denoiser behaves as a true posterior-mean estimator, Theorem 1 states that the reverse-diffusion output satisfies E[Ŷ_{t+1:t+k}|H_

Load-bearing premise

The load-bearing premise is that the Transformer's prediction errors on smoothed rainfall look statistically the same as the Gaussian noise the diffusion model was trained to remove; if they do not, the denoiser will remove signal along with noise and pull forecasts toward the average rainfall pattern.

Editorial extensions

If this is right

  • On South Australian precipitation data, the full pipeline beats the Non-stationary Transformer baseline by up to 56.7% MSE and 21.1% MAE, with gains at every forecast horizon from 24 to 720 days.
  • Ablations show the two smoothing/denoising stages are the load-bearing parts: removing the Gaussian noise injection raises MSE by 72.7%, and removing the diffusion denoising raises MSE by 30.9%.
  • The best noise level is a broad plateau around αnoise = 0.1–0.2, so the method does not require fine tuning of that parameter.
  • The framework is modular: a standard Transformer can replace the Non-stationary Transformer with only a 14.2% MSE penalty, and reducing diffusion steps from 1000 to 100 costs 10.9% MSE.

Reading between the lines

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

  • The recipe should transfer beyond rainfall: any zero-heavy, right-skewed target—insurance losses, disease counts, species abundance—has the same point-mass-at-zero structure, and the Gaussian perturbation step does not use any rainfall-specific knowledge.
  • The theoretical guarantee is only as good as the Gaussian-error assumption; if a real predictive model's errors are heavy-tailed or skewed, the denoiser could partly erase signal, and the reported MSE gains may partly come from shrinkage toward the marginal distribution rather than from recovering true conditional means.
  • A natural extension is to let the diffusion model output full predictive distributions instead of point forecasts, turning the framework into a probabilistic forecaster well suited to extreme-event and uncertainty-aware applications.
  • Because the paper reports degraded performance when the zero proportion exceeds 90%, an adaptive noise level tied to the local zero ratio could extend the framework into the most extreme regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes ZIDF, a framework for zero-inflated precipitation forecasting. The method adds Gaussian noise to the zero-inflated target to obtain a smoothed target, trains a Transformer forecaster on that smoothed target with an MSE loss, and then applies a DDPM reverse process to the forecaster's predictions, claiming the denoised output recovers the original clean distribution. The central theoretical result, Theorem 1 (Eq. 21), asserts that the denoised predictions equal the conditional mean of the clean target given the history. The empirical sections report synthetic simulations and a South Australia case study, with MSE/MAE comparisons against several Transformer and zero-inflated baselines, plus an ablation study.

Significance. If the central theorem were valid, ZIDF would offer a generic and practical wrapper for adapting arbitrary forecasters to zero-inflated targets, a problem of real importance in precipitation and other sparse time-series domains. The proposed pipeline is simple, the paper includes experiments across many horizons, and the authors provide a public repository. However, the theoretical guarantee is not established: the DDPM denoiser is trained on forward-diffused clean samples, and there is no justification for treating Transformer predictions as noisy observations of clean targets at a particular diffusion step. The reported gains, especially the ablation in Table 2, are consistent with the denoiser acting as a shrinkage estimator toward the marginal rainfall distribution rather than as a conditional-mean corrector. As presented, the contribution is an empirical recipe with a plausible intuition, not a validated principled method.

major comments (4)
  1. [Section 3.4, Eq. (21) and proof sketch] The proof of Theorem 1 is invalid. DDPM training (Eq. 18) makes the denoiser approximate the posterior mean E[Y(0)|Y(s)] under the forward diffusion of clean samples. At inference, Y(τ) is constructed from the Transformer prediction (or, as written, from the future smoothed target Y'_{t+1:t+k}) plus independent Gaussian noise. This input is not a forward-diffused clean sample, so the statement 'the reverse denoiser then outputs E[Y_{t+1:t+k} | Y(τ), \bar H_t]' does not follow from the DDPM training objective. In fact, the DDPM posterior mean for such an input is a shrinkage toward the marginal mean of Y, not E[Y|H_t]. The proof also embeds Y'_{t+1:t+k} (the future smoothed target) rather than the H_t-measurable prediction Ŷ', and the step claiming conditioning on \bar H_t is equivalent to conditioning on H_t is not justified. Thus Eq. (21) is unsupported.
  2. [Section 3.3] The assertion that the noise level in the predictions 'naturally matches' the DDPM high-noise state is not derived. The Gaussian perturbation in Eq. (7)/(22) has scale α_noise·μ_Y, while the diffusion forward process in Eq. (13) uses scales √ᾱ_s and (1−ᾱ_s). No relation between α_noise and a diffusion step τ is given, so the inference-time input distribution may be far from the training-time noisy inputs. This distribution shift is a likely source of systematic bias in the denoised output.
  3. [Section 4.3 / Table 2] The ablation result (w/o diffusion: MSE 0.5724 vs full: 0.3956) is consistent with a shrinkage effect: in a zero-inflated target, shrinking predictions toward the marginal mean can reduce MSE without improving the conditional mean. To support the claimed mechanism and the abstract's 56.7% MSE reduction, the paper needs distributional metrics (e.g., CRPS, quantile/occurrence accuracy, reliability diagrams) and statistical uncertainty (multiple seeds, confidence intervals). The single point estimates in Table 1 do not establish that the denoising step is doing what Theorem 1 claims.
  4. [Section 3.4, Assumption 1 (Eq. 20)] Assumption 1 asserts that the Transformer trained on smoothed targets is exactly unbiased in the conditional-mean sense. This is a strong property that does not follow from MSE minimization (Eq. 11), especially with finite data and a zero-inflated target. Since the theorem's conclusion depends on it, the assumption should be empirically validated (e.g., calibration on held-out data) or the theorem should be reframed as a heuristic statement.
minor comments (5)
  1. [Section 2.2, Eq. (5)] The claimed monotonic decrease of non-zero attention with π0 assumes unspecified distributions for similarity scores of zero and non-zero keys. Without these assumptions the formula is not a rigorous proof; please label the analysis as heuristic.
  2. [Section 3.4 proof sketch] The notation alternates between Y'_{t+1:t+k} and Ŷ'_{t+1:t+k}; clarify which quantity is embedded into the forward process.
  3. [Section 3.3] The statement E[||Ŷ'_{t+1:t+k} − Ŷ_{t+1:t+k}||^2] → 0 lacks a limit argument; either provide a precise convergence statement or remove it.
  4. [Conclusion] Limitations are listed but not quantified (e.g., 'performance drop under extreme zeros (>90%)' and 'longer inference time ≈3×'). Please provide numbers if these statements are retained.
  5. [Throughout] There are grammatical and formatting issues, including 'shows that attention can fail under non-stationary time series because the similarity matrix drives' in Section 2.2 and 'the github1' reference format. A careful proofread is needed.

Circularity Check

1 steps flagged · score 4.0 of 10

Theorem 1's equality is assumed through the denoiser's conditional-mean capability; the empirical benchmark comparisons remain independent.

  1. self definitional [Section 3.4 (Theorem 1, Eq. 21) and proof sketch]
    "Under Assumption 1 and the denoising capability of the diffusion model, the final predictions after reverse diffusion satisfy: E[ ˆYt+1:t+k | Ht] = E[Yt+1:t+k | Ht]. ... From the standard DDPM identity under the simple loss, the reverse denoiser then outputs ˆYt+1:t+k = E[ Yt+1:t+k | Y (τ ), ¯Ht ], which represents the posterior mean of the clean target given the noisy input."

    The 'standard DDPM identity' applies to a denoiser trained on forward-diffused clean samples, i.e. Y(s)=√ᾱ_sY(0)+√(1−ᾱ_s)ε. At inference, Y(τ) is built from the Transformer's smoothed-target prediction plus independent noise, not from a clean target. So the identity can be invoked only if the denoiser is already assumed to map any noisy prediction to E[Y_{t+1:t+k}|Y(τ),H_t], which is precisely what Theorem 1 asserts. Assumption 1 (Eq. 20) only gives f as the conditional mean of the smoothed Y', not of the clean Y; the proof inserts the missing link as 'the denoising capability' and as the DDPM identity. Thus Eq. 21 is an assumption restated as a theorem rather than a derivation.

full rationale

The paper's empirical comparisons (Table 1, ablation study) are independent external benchmarks and do not reduce to fitted parameters, so the main experimental contribution is not circular. However, the central theoretical guarantee, Theorem 1, is self-definitional: it assumes the denoiser computes E[Y|Y(τ),H_t] for inputs that are not generated by the diffusion forward process on clean targets. That assumption is exactly the content of the theorem, so the derivation chain contains a circular step. Because the theoretical claim is load-bearing only as a stated guarantee and the empirical results stand on their own, the circularity is moderate rather than total.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework introduces one tuned hyperparameter (alpha_noise) and relies on a diffusion step count as a tunable. The central theorem rests on two assumptions that are stated or unstated but not verified. No new physical entities, particles, or forces are introduced.

free parameters (2)
  • alpha_noise = 0.1
    Noise injection ratio sigma = alpha_noise * mu_Y. Tuned on synthetic data (optimal 0.1-0.2) and fixed at 0.1 for real-world experiments without a dedicated validation split (Section 4.1.2, 4.2).
  • diffusion steps S = 1000
    Total DDPM steps; standard choice from Ho et al. (2020), but ablation shows S=100 yields 10.9% higher MSE, so S=1000 is a tuned hyperparameter (Section 4.3).
assumptions (5)
  • ad hoc to paper Assumption 1: The predictive component f trained on smoothed targets is unbiased, i.e., E[Y'_{t+1:t+k} | H_t] = f(Y'_{t-h+1:t}, X_{t-h+1:t}).
    Stated as Assumption 1 (Eq. 20) without empirical verification. An MSE-trained Transformer is not guaranteed to be conditionally unbiased, especially with non-Gaussian smoothed targets.
  • ad hoc to paper The DDPM reverse process gives the exact posterior mean E[Y | Y(tau)] for inputs that are Transformer predictions embedded into the forward process.
    The paper relies on this in Theorem 1's proof sketch. DDPM training on forward-diffused clean samples does not guarantee correct denoising of arbitrary model predictions, whose error distribution is not the forward diffusion noise.
  • standard math Gaussian convolution of a Dirac mixture yields an infinitely differentiable density (Lemma 1).
    Standard property of convolution with a Gaussian kernel, cited to Rudin and Folland. This supports the smoothing step and is not in dispute.
  • standard math The DDPM training objective in Eq. 18 simplifies to a mean-squared error between true and predicted noise.
    Standard DDPM result (Ho et al. 2020). Used to justify the denoiser's training.
  • domain assumption In Section 2.2, the similarity scores S_ij are treated as fixed while varying the zero-inflation ratio pi_0, so that total attention mass on non-zeros decreases monotonically with pi_0.
    The analysis ignores that learned embeddings adapt to pi_0. In a trained model, S_ij for zero and non-zero positions are not independent of the data mix, so the monotonicity claim is not generally valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Noise to Precision: A Diffusion-Driven Approach to Zero-Inflated Precipitation Prediction." pith.science (2026). https://pith.science/paper/4TOU7SHM

@misc{pith2026250910501,
  author       = {Pith},
  title        = {Pith review of: From Noise to Precision: A Diffusion-Driven Approach to Zero-Inflated Precipitation Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TOU7SHM}},
  note         = {Machine review of arXiv:2509.10501}
}
read the original abstract

Zero-inflated data pose significant challenges in precipitation forecasting due to the predominance of zeros with sparse non-zero events. To address this, we propose the Zero Inflation Diffusion Framework (ZIDF), which integrates Gaussian perturbation for smoothing zero-inflated distributions, Transformer-based prediction for capturing temporal patterns, and diffusion-based denoising to restore the original data structure. In our experiments, we use observational precipitation data collected from South Australia along with synthetically generated zero-inflated data. Results show that ZIDF demonstrates significant performance improvements over multiple state-of-the-art precipitation forecasting models, achieving up to 56.7\% reduction in MSE and 21.1\% reduction in MAE relative to the baseline Non-stationary Transformer. These findings highlight ZIDF's ability to robustly handle sparse time series data and suggest its potential generalizability to other domains where zero inflation is a key challenge.

Figures

Figures reproduced from arXiv: 2509.10501 by the authors.

Figure 1
Figure 1. Challenges of Zero-Inflated Drought Data: Visualizing Global Extreme Drought Trends and Sparse Precipitation (prate) Events. structural zeros (e.g., truly dry days) from non-structural zeros (ze￾ros that occasionally appear within the main precipitation distribu￾tion) by modeling zeros via a Bernoulli process and non-zeros via a continuous distribution [23]. In the statistical and actuarial literature, more sophisti… view at source ↗
Figure 2
Figure 2. The overall architecture of the Zero-Inflated Diffusion Framework (ZIDF) for modeling sparse precipitation events. The process starts by adding Gaussian noise to the zero-inflated target Y to generate a smoothed distribution Y ′ . Using historical features X, the prediction model forecasts noisy future values Yˆ ′ . The diffusion process models the forward noise distribution, while the denoising process recovers the… view at source ↗
Figure 3
Figure 3. Architecture of Non-stationary Transformer [14]. The encoder and decoder of the Non-stationary Transformer con￾sist of 2 and 1 layers, respectively. Following the default settings from [14], the model has a hidden dimension of 512, 8 attention heads and a feedforward network dimension of 2048. The mov￾ing average window size is set to 25, with an attention factor of 1. GELU activation is used, and the encoder incorp… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The figure illustrates the performance of different models under varying zero inflation proportions and noise levels. The left two panels display non-zero MAE and non-zero MSE across different zero proportions, showing that the ZIDF model consistently outperforms other…
Figure 5
Figure 5. Figure 5: Study Area: South Australia (34°S–38°S, 129°E–141°E). The state is characterized by semi-arid to arid climate conditions with highly variable precipitation patterns. We extracted and converted climate data from 1948 to 2014 for the South Australia region using the NCEP…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages

  1. [1]

    K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian. Accurate medium-range global weather forecasting with 3D neural networks.Na- ture, 619(7970):533–538, 2023

    work page 2023

  2. [2]

    Chen, C.-L

    S.-A. Chen, C.-L. Li, N. Yoder, S. Ö. Arik, and T. Pfister. TSMixer: An all-MLP architecture for time series forecasting. Transactions on Machine Learning Research, 2023

    work page 2023

  3. [3]

    P. K. Dunn and G. K. Smyth. Series evaluation of Tweedie exponential dispersion model densities. Statistics and Computing, 15(4):267–280, 2005

    work page 2005

  4. [4]

    C. Feng, L. Li, and C. Xu. Advancements in predicting and model- ing rare event outcomes for enhanced decision-making. BMC Medical Research Methodology, 23(1):243, 2023

    work page 2023

  5. [5]

    C. X. Feng. A comparison of zero-inflated and hurdle models for mod- eling zero-inflated count data. Journal of Statistical Distributions and Applications, 8(1):8, 2021

    work page 2021

  6. [6]

    G. B. Folland. Real Analysis: Modern Techniques and Their Applica- tions. Wiley, 2nd edition, 1999

    work page 1999

  7. [7]

    Hegde, M

    P. Hegde, M. Heinonen, and S. Kaski. Variational zero-inflated Gaus- sian processes with sparse kernels. In Conference on Uncertainty in Artificial Intelligence (UAI), 2018

    work page 2018

  8. [8]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020

    2020

Show all 35 references
  1. [9]

    Kalnay, M

    E. Kalnay, M. Kanamitsu, R. Kistler, W. Collins, D. Deaven, L. Gandin, M. Iredell, S. Saha, G. White, J. Woollen, Y . Zhu, A. Leetmaa, R. Reynolds, M. Chelliah, W. Ebisuzaki, W. Higgins, J. Janowiak, K. C. Mo, C. Ropelewski, J. Wang, R. Jenne, and D. Joseph. The NCEP/NCAR 40-y...

    1996

  2. [10]

    S. Kong, J. Bai, J. H. Lee, D. Chen, A. Allyn, M. Stuart, M. Pinsky, K. Mills, and C. P. Gomes. Deep hurdle networks for zero-inflated multi-target regression: Application to multiple species abundance esti- mation. arXiv preprint arXiv:2010.16040, 2020

    2010 arXiv

  3. [11]

    D. Lambert. Zero-inflated Poisson regression, with an application to defects in manufacturing. Technometrics, 34(1):1–14, 1992

    1992

  4. [12]

    Lindeberg

    T. Lindeberg. Discrete approximations of Gaussian smoothing and Gaussian derivatives. Journal of Mathematical Imaging and Vision, 66 (5):759–800, 2024

    2024

  5. [13]

    W. Ling, B. Cheng, Y . Wei, J. Z. Willey, and Y . K. Cheung. Statistical inference in quantile regression for zero-inflated outcomes. Statistica Sinica, 32(3):1411, 2022

    2022

  6. [14]

    Y . Liu, H. Wu, J. Wang, and M. Long. Non-stationary transformers: Exploring the stationarity in time series forecasting. In Advances in Neural Information Processing Systems, 2022

    2022

  7. [15]

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long. iTrans- former: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625, 2023

    2023 arXiv

  8. [16]

    J. Mullahy. Specification and testing of some modified count data mod- els. Journal of Econometrics, 33(3):341–365, 1986

    1986

  9. [17]

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A time se- ries is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730, 2023

    2023 arXiv

  10. [18]

    Ravuri, K

    S. Ravuri, K. Lenc, M. Willson, D. Kangin, R. Lam, P. Mirowski, M. Fitzsimons, M. Athanassiadou, S. Kashem, S. Madge, R. Prudden, A. Mandhane, A. Clark, A. Brock, K. Simonyan, R. Hadsell, N. Robin- son, E. Clancy, A. Arribas, and S. Mohamed. Skillful precipitation now- casting...

    2021

  11. [19]

    R. Rew, E. Hartnett, and J. Caron. NetCDF-4: Software implementing an enhanced data model for the geosciences. In Proceedings of the 22nd International Conference on Interactive Information Processing Systems for Meteorology, Oceanography, and Hydrology, 2006

    2006

  12. [20]

    C. E. Rose, S. W. Martin, K. A. Wannemuehler, and B. D. Plikaytis. On the use of zero-inflated and hurdle models for modeling vaccine adverse event count data. Journal of Biopharmaceutical Statistics , 16(4):463– 481, 2006

    2006

  13. [21]

    W. Rudin. Real and Complex Analysis . McGraw–Hill, 3rd edition, 1987

    1987

  14. [22]

    Shen and J

    L. Shen and J. Kwok. Non-autoregressive conditional diffusion models for time series prediction. arXiv preprint arXiv:2306.05043, 2023

    2023 arXiv

  15. [23]

    Sheng, S

    Z. Sheng, S. Wen, Z.-k. Feng, J. Gong, K. Shi, Z. Guo, Y . Yang, and T. Huang. A survey on data-driven runoff forecasting models based on neural networks. IEEE Transactions on Emerging Topics in Computa- tional Intelligence, 7(4):1083–1097, 2023

    2023

  16. [24]

    E. M. Stein and R. Shakarchi. Fourier Analysis: An Introduction . Princeton University Press, 2003

    2003

  17. [25]

    K. E. Trenberth and J. M. Caron. The southern oscillation revisited: Sea level pressures, surface temperatures, and precipitation. Journal of Climate, 13(3):435–455, 2000

    2000

  18. [26]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017

    2017

  19. [27]

    S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y . Zhang, and J. Zhou. TimeMixer: Decomposable multiscale mixing for time series forecast- ing. arXiv preprint arXiv:2405.14616, 2024

    2024 arXiv

  20. [28]

    Wilson, A

    T. Wilson, A. McDonald, A. H. Galib, P.-N. Tan, and L. Luo. Beyond point prediction: Capturing zero-inflated & heavy-tailed spatiotemporal data with deep extreme mixture models. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2020...

    2020

  21. [29]

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long. TimesNet: Temporal 2D-variation modeling for general time series analysis.arXiv preprint arXiv:2210.02186, 2023

    2023 arXiv

  22. [30]

    Y . Yang, W. Qian, and H. Zou. Insurance premium prediction via gradi- ent tree-boosted Tweedie compound Poisson models. Journal of Busi- ness & Economic Statistics, 36(3):456–470, 2018

    2018

  23. [31]

    D. Yoon, M. Seo, D. Kim, Y . Choi, and D. Cho. Probabilistic weather forecasting with deterministic guidance-based diffusion model. In Eu- ropean Conference on Computer Vision (ECCV), pages 108–124, 2024

    2024

  24. [32]

    D. S. Young, E. S. Roemmele, and P. Yeh. Zero-inflated modeling part I: Traditional zero-inflated count regression models, their applications, and computational tools. WIREs Computational Statistics, 14(1):e1541, 2022

    2022

  25. [33]

    S. Zhai, T. Likhomanenko, E. Littwin, D. Busbridge, J. Ramapuram, Y . Zhang, J. Gu, and J. M. Susskind. Stabilizing transformer training by preventing attention entropy collapse. In International Conference on Machine Learning, pages 40770–40803, 2023

    2023

  26. [34]

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. arXiv preprint arXiv:2012.07436, 2021

    2012 arXiv

  27. [35]

    H. Zhou, W. Qian, and Y . Yang. Tweedie gradient boosting for ex- tremely unbalanced zero-inflated data. Communications in Statistics— Simulation and Computation, 51(9):5507–5529, 2022

    2022

Pith tools

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