- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
Summary:
The authors argue that historical observations inherently contain noise, which may cause forecasting models to absorb spurious and non-generalizable relations. To address this issue, they propose FenDiM, a framework whose core integrates diffusion-based rectification, frequency-domain compensation, and Mamba-based temporal modeling.
Critical strengths
(1) Overall, the problem addressed in this paper is clear.
(2) The paper is well-structured.
(3) The empirical results demonstrate that the proposed FenDiM consistently achieves strong forecasting performance across varying proportions of missing observations.
Critical weaknesses
Reasons for Rejection:
I lean towards rejecting this paper due to the following concerns:
(1) The rationale for employing a diffusion-based approach for dependency rectification remains inadequately justified. It is unclear why diffusion is strictly necessary.
(2) The introduction of Frequency-Compensated Temporal Learning feels disjointed. Specifically, the claim that “local time-domain evidence is unreliable” (Line 68) lacks proper context and a logical transition, which may confuse readers.
(3) The Mamba-based forecasting backbone offers no substantial innovation. The proposed framework reads more like a straightforward patchwork of existing techniques rather than a fundamentally novel contribution, thereby severely limiting its technical novelty.
(4) The numerical experiments are restricted to low-dimensional datasets (e.g., the Weather dataset, which contains only 21 variables). The paper lacks rigorous comparative evaluations on standard, higher-dimensional multivariate time-series benchmarks, such as the electricity and traffic datasets, raising concerns about the scalability of the proposed method.
Things to improve the paper that did not impact the score:
(1) The title of the submitted manuscript is inconsistent with the one provided in the open-source code repository; please unify them.
(2) In Line 145 (“where the target remains clean”), the symbol does not actually appear in Equation (2). Please correct this notation mismatch.
(3) The overall methodology diagram (Figure 1) is overly cluttered, making it difficult for readers to grasp the core insights. I suggest the authors simplify the workflow diagram to better highlight the main technical contributions.
Questions for authors
if the objective of the reverse diffusion process is merely to reconstruct the original , wouldn’t the model simply learn to recover the corrupted as is? How can it possibly “rectify” the representation and eliminate misleading dependencies under such an objective?
4 comments
I thank the authors for their response.
It appears that the authors may have forgotten to upload the updated PDF, as I do not see Section 5.7 or Table 16 in the current version. Nevertheless, the authors' rebuttal has addressed most of my previous concerns. However, to be honest, the current quality of the manuscript still falls short of the acceptance bar for ICLR in terms of novelty, clarity of presentation, and the sufficiency of experimental validation. I hope the authors will not be discouraged, and I encourage you to further refine and strengthen the paper. I wish you all the best in your future research!
In fact, I would like to offer a piece of research advice: the method proposed in your paper incorporates too many elements, involving numerous modular components and loss terms. Generally, identifying a single, genuine problem from the existing literature and proposing a novel solution to address it is a better approach to conducting research. Having gone through a similar stage myself, I wanted to share my perspective, and I hope you find it helpful.
Additional Results on High-Dimensional Data, Electricity and Traffic (Section 5.7)
We thank the reviewer again. We report here the high-dimensional results requested in W4, which form Section 5.7 (Table 16) of our revised manuscript. Setting: L = 96; Standard is averaged over T ∈ {96, 192, 336, 720}; 50% missingness and Gaussian noise use T = 192, following the protocol in App. B.3. Entries are MAE / MSE; best in bold.
| Data | Method | Params (M) | Mem. (GB) | Time (s/epoch) | Standard | Missing 50% | Gaussian noise |
|---|---|---|---|---|---|---|---|
| Electricity (C=321) | FenDiM | 24.86 | 22.8 | 197 | 0.274 / 0.180 | 0.325 / 0.222 | 0.277 / 0.182 |
| FenDiM w/o PDR | 24.30 | 9.6 | 124 | 0.360 / 0.272 | 0.369 / 0.284 | 0.362 / 0.273 | |
| iTransformer | 4.83 | 27.6 | 254 | 0.366 / 0.284 | 0.377 / 0.297 | 0.366 / 0.284 | |
| xPatch | 0.14 | 4.5 | 84 | 0.293 / 0.206 | 0.352 / 0.253 | 0.295 / 0.207 | |
| Traffic (C=862) | FenDiM | 63.03 | 19.3 | 667 | 0.339 / 0.480 | 0.361 / 0.519 | 0.341 / 0.481 |
| FenDiM w/o PDR | 60.97 | 13.6 | 276 | 0.349 / 0.648 | 0.354 / 0.656 | 0.349 / 0.649 | |
| iTransformer | 6.41 | 163.6 | 291 | 0.344 / 0.642 | 0.349 / 0.648 | 0.344 / 0.642 | |
| xPatch | 0.15 | 13.9 | 196 | 0.394 / 0.604 | 0.440 / 0.681 | 0.396 / 0.606 |
FenDiM attains the lowest error in 11 of the 12 entries. Its MSE is 12.1–12.6% lower than that of the strongest baseline on Electricity (xPatch) and 19.9–20.6% lower than the best baseline in each setting on Traffic, where its MAE stays close to iTransformer's and is higher only under 50% missingness. Its peak memory does not grow from Electricity to Traffic (22.8 and 19.3 GB); the main cost is training time on Traffic (667 vs. 291 s/epoch for iTransformer).
The table also shows where the gain comes from. Without PDR, FenDiM is on par with iTransformer (MSE 0.272 vs. 0.284 on Electricity, 0.648 vs. 0.642 on Traffic), and adding PDR lowers MSE by 21.8–33.8% on Electricity and 20.9–25.9% on Traffic. The improvement on these datasets therefore comes from PDR rather than from the backbone, consistent with the ablation in Fig. 2.
We hope these results address your concern about scalability to high-dimensional data.
Thank You for the Advice and Clarify the Novelty of FenDiM
Thank you very much for your candid reply and for taking the time to share your research advice; we sincerely appreciate it and take it to heart.
We fully agree with your advice: research should start from one genuine problem and solve it with one clear idea. FenDiM is built on exactly this principle; our original presentation obscured it by describing every module and loss term with equal weight. We would like to state the problem and the idea plainly.
One problem. Missing values and noise do not only perturb individual values; they alter the temporal dependencies that a forecaster learns from the observed history. Existing work follows two routes. It either repairs the data first (impute or denoise, then forecast), which optimizes reconstruction rather than forecasting and passes reconstruction errors to the forecaster, or it builds stronger forecasters, which fit the observed series more closely and therefore also fit its corrupted dependencies more closely. Neither route corrects the dependencies themselves, which is where corruption actually harms forecasting.
One idea. FenDiM changes both what is corrected and what the correction is optimized for: it rectifies the learned dependency representations rather than the values, and it does so for forecasting accuracy on the clean future rather than for reconstruction fidelity. This is PDR, our main contribution. To our knowledge, diffusion has not previously been used this way: as an end-to-end, prediction-oriented rectifier inside the forecaster, trained in a single stage, rather than as a generator of values (TimeGrad, CSDI, SSSD) or a separate reconstruction stage. Diffusion is also not an arbitrary choice. Corruption severity is unknown and varies across samples, so the rectifier must correct deviations of unknown size; a denoiser trained across noise levels removes them step by step, which is exactly the progressive rectification PDR is named for.
One design, not a patchwork. Mamba and FCTL follow from this single choice rather than adding new ideas. Progressive rectification is iterative, so we implement the denoiser with Mamba to keep it affordable (38.5% less total inference time than an attention denoiser, App. E). Missingness and noise act on individual time steps, and rectifying dependencies does not fully undo such local damage, so FCTL recalibrates the rectified representation with global frequency evidence through a gated residual rather than a separate branch. One forecasting-driven objective trains all three.
In short, FenDiM's contribution is a new formulation of robust forecasting, from observation-driven dependency extraction to prediction-oriented dependency rectification, rather than a combination of modules. We agree that our original manuscript did not make this hierarchy clear, and we will state it explicitly in the revision.
Thank you again for your encouragement and for sharing your experience; it has genuinely improved how we present this work. We hope these clarifications make the single problem and its solution clearer, and would be grateful if you would take them into account in your final assessment.
Clarifying End-to-End Dependency Rectification: Why Diffusion, How Frequency Compensation Helps, and Mamba’s Role
Thank you for your careful reading and constructive comments. Key clarification: FenDiM jointly trains latent refinement and forecasting end to end, with clean-future supervision.
W1. Why diffusion?
We have clarified the design rationale in Sections 1 and 4.2. Diffusion provides noise-level-conditioned refinement: a shared denoiser is trained across 1,000 diffusion levels and conditioned on available observations. This motivates conditional refinement without claiming mathematical necessity or equating diffusion noise with observation corruption.
PDR operates on latent representations, and its output feeds the forecaster. The encoder, refiner, and predictor are jointly optimized, without a separately pretrained or frozen imputation stage. Sections 4.2–4.4 now make this distinction explicit. Removing PDR increases error on all four ablation datasets (Figure 2), supporting its empirical contribution.
Q1. Why does reverse diffusion not simply recover corrupted Z0?
We have clarified the two supervision signals in Sections 4.2 and 4.4, and depicted their paths in Figure 1. Eq. (17) predicts artificial diffusion noise; Eq. (19) combines this auxiliary objective with clean-future prediction and spectral supervision. The prediction loss Lpred backpropagates through the differentiable refinement path into the denoiser and encoder. Consequently, refinement is optimized for predicting the clean future, rather than solely reconstructing its historical input.
The distinction is between the auxiliary denoising target and the complete forecasting objective. Conditional denoising supplies refinement conditioned on partial evidence; clean-future supervision directs that refinement toward forecasting utility. We have separated these roles after Eq. (19), rather than treating reconstruction of Z0 as proof of dependency rectification.
The auxiliary weight is small (λpdr=0.005). Appendix D.1/Figure 3 reports average ETTh1 MAE/MSE reductions of 1.14%/1.12% relative to zero diffusion weight. Together with Figure 2, this supports PDR's empirical utility.
W2. FCTL motivation (Line 68).
We have rewritten the transition in the Introduction and the opening of Section 4.3: missing entries and measurement noise perturb local historical evidence; residual distortions can remain after refinement; sequence-level magnitude and phase descriptors provide complementary context for temporal prediction. Eqs. (10)–(14) specify how these descriptors modulate the temporal carrier through a gated residual.
Section 4.3 now explains the complementary roles of magnitude (spectral energy) and phase (temporal alignment). We have clarified in Appendix F that Eq. (33) bounds the specified band-pooling construction under its assumptions; it does not establish a bound for the full phase-aware module. Figure 2 reports the effect of removing FCTL; Table 12 separately evaluates phase-aware versus magnitude-only conditioning.
W3. Novelty and Mamba.
We have revised the contribution statement in Section 1 and the discussion in Section 5.5 to identify the contribution precisely: prediction-oriented conditional latent refinement coupled with spectral modulation for robust forecasting. PDR produces the representation from which FCTL derives its conditioning descriptors, and the joint objective trains their forecasting interaction. Mamba is an established computational component, not a claimed new primitive.
Appendix E/Figure 4 documents its efficiency role:
Table 11 further compares rectifier backbones. These results document Mamba's computational role.
W4. Higher-dimensional evaluation.
We have added Electricity (321 variables) and Traffic (862 variables) evaluations in Section 5.7, with complete results in Table 16 and settings in Appendix H.4. The comparison includes FenDiM, no PDR, iTransformer, and xPatch under matched evaluation conditions, reporting MAE/MSE, parameter counts, peak memory, and complete inference latency. We distinguish sequence-length efficiency from scalability in variable count and assess the latter through these additional benchmarks.
Minor comments.
We have (1) aligned the repository title with the manuscript; (2) revised the sentence following Eq. (2) to refer explicitly to the clean future target Y defined in Section 3; and (3) simplified Figure 1 to highlight refinement, spectral modulation, prediction, and supervision, moving internal details to the appendix.
We would appreciate your reassessment in light of these clarifications and additional evidence.