acceptodds
Under review as a conference paper at ICLR 2027

Better Dependencies, Better Forecasts: Dependency Rectification for Robust Time Series Forecasting

Abstract

Time series forecasting typically assumes that predictive dependencies can be reliably learned from observed histories. However, real-world observations are often imperfect: noise, missing values, and localized distortions not only perturb individual values but also alter the apparent temporal dependencies that models rely on. As a result, directly fitting observed correlations may cause forecasting models to absorb spurious and non-generalizable relations, degrading robustness under corrupted inputs. To address this issue, we propose FenDiM, a Frequency-Enhanced Diffusion-Mamba forecasting architecture for prediction-oriented dependency rectification. FenDiM progressively suppresses corruption-induced dependencies through diffusion-based rectification, uses frequency-domain evidence to compensate for unreliable time-domain patterns, and leverages Mamba for efficient long-range temporal modeling. Across six benchmarks, FenDiM achieves the best horizon-averaged MSE on five datasets under standard observations, reducing MSE by 20.7% on average relative to the strongest competing baseline across these five datasets. At a prediction length of 192, it ranks within the top two in both MAE and MSE across all tested dataset–missing-rate settings and remains competitive under Gaussian noise. These results demonstrate the effectiveness of prediction-oriented dependency rectification for robust time series forecasting. The code and datasets are available at https://anonymous.4open.science/r/FenDim-5035.

Primary area:
learning on time series and dynamical systems
Keywords:
Robust time series forecasting, Multivariate time series forecasting, Imperfect observations, Diffusion models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pNfJOpenReview ↗
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 Y\text{Y} remains clean”), the symbol Y\text{Y} 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 Z0Z_0, wouldn’t the model simply learn to recover the corrupted Z0Z_0 as is? How can it possibly “rectify” the representation and eliminate misleading dependencies under such an objective?

4 comments
AuthorsOpenReview ↗

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 Z0Z_0?

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 LpredL_{\mathrm{pred}} 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 Z0Z_0 as proof of dependency rectification.

The auxiliary weight is small (λpdr=0.005\lambda_{\mathrm{pdr}}=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:

Per-batch forward timeAttention denoiserMamba denoiser
Diffusion rectification0.324 s0.153 s
Whole model0.397 s0.182 s

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 YY 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.

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.

AuthorsOpenReview ↗

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.

DataMethodParams (M)Mem. (GB)Time (s/epoch)StandardMissing 50%Gaussian noise
Electricity (C=321)FenDiM24.8622.81970.274 / 0.1800.325 / 0.2220.277 / 0.182
FenDiM w/o PDR24.309.61240.360 / 0.2720.369 / 0.2840.362 / 0.273
iTransformer4.8327.62540.366 / 0.2840.377 / 0.2970.366 / 0.284
xPatch0.144.5840.293 / 0.2060.352 / 0.2530.295 / 0.207
Traffic (C=862)FenDiM63.0319.36670.339 / 0.4800.361 / 0.5190.341 / 0.481
FenDiM w/o PDR60.9713.62760.349 / 0.6480.354 / 0.6560.349 / 0.649
iTransformer6.41163.62910.344 / 0.6420.349 / 0.6480.344 / 0.642
xPatch0.1513.91960.394 / 0.6040.440 / 0.6810.396 / 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.

AuthorsOpenReview ↗

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.

Reviewer_gxWiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that corrupted observations (missing values, noise) harm forecasting not only by perturbing values but by inducing spurious temporal dependencies, and proposes "prediction-oriented dependency rectification". The model, FenDiM, combines Progressive Dependency Rectification (PDR), a latent conditional diffusion process with a Mamba denoiser and 2 inference steps, and Frequency-Compensated Temporal Learning (FCTL), which uses log-magnitude and sin/cos phase descriptors for FiLM-style gating, on a Mamba backbone trained with MSE, diffusion and spectral L1 losses. Experiments cover six LTSF benchmarks with L=96 under clean inputs (Table 1), 30/50/70% missingness (Tables 2, 5-7), Gaussian noise (Table 8) and missing+noise (Table 9) against 12 retrained baselines, plus ablations (Fig. 2, Tables 11-12), efficiency (Table 10) and a Lipschitz-style analysis (App. F). The headline claim is a 20.7% average MSE reduction over the strongest baseline on five of six clean datasets.

Critical strengths
  1. Robust forecasting under imperfect inputs is a practically relevant problem, and the paper specifies an explicit corruption model (Eq. 1, App. B.3).
  2. The robustness evaluation is broad, covering several missing rates, Gaussian noise and compound corruption on six datasets with full per-horizon tables (Tables 5-9); if the numbers were reliable, this protocol would be a useful template.
Critical weaknesses
  1. The core "dependency rectification" claim is not supported by evidence specific to it. Seems no experiment defines or measures a dependency or compares ||Z_rect - Z*|| with ||Z0 - Z*||;

  2. The results cannot currently be trusted or compared with the literature. Clean L=96 baselines are far from published values (iTransformer ETTh1 MAE/MSE 0.212/1.283 vs. 0.447/0.454), and MSE is 15-40x MAE² for almost all models, with evaluation scale, C_tar and masked positions unspecified. Why this evaluation protocol is right suitable for your research?

Questions for authors
  1. Please provide direct evidence that dependencies are actually rectified, e.g., final test errors of w/o PDR under missingness and noise.
  2. Please clarify the evaluation scale and explain the identical values across different models and settings in Tables 1 and 5-9.
2 comments
AuthorsOpenReview ↗

Dependency Rectification and Evaluation Reliability: Direct Evidence and Numerical Audit

Thank you for your careful review. We have expanded the evaluation protocol, numerical audit, and PDR analyses.

1. Forecasting task, target variables, and evaluation protocol. The task in Section 3 is forecasting future targets from possibly corrupted histories. Appendix B.1, Eqs. (21)-(22) evaluates future test windows, prediction steps, and target variables. Evaluating only masked historical positions would instead measure imputation, a different task.

We have consolidated the protocol in Section 5.1, Appendices B.1-B.3, and Appendix H.1, Table 15(a). It records chronological splits, M/MS mode, target indices, scaler fitting range, evaluation scale, inverse transformation, training/test corruption, and scored tensor shapes for each implementation.

For the all-variable task in Appendix B.2 and Table 3, the target-channel counts are 7 for ETT, 21 for Weather, and 8 for Exchange. The six learned index features are internal features and are excluded from the physical targets and metric denominator. Historical corruption leaves future targets unchanged.

We have added Eq. (43) to specify that both metrics use the same residual tensor after identical target selection and evaluation-scale transformation. MAE averages absolute residuals, while MSE averages squared residuals, over all test windows, forecast steps, and target variables. Each run is evaluated on the full test set before averaging across runs; four-horizon averages are distinguished from single-horizon results.

2. The MSE/MAE-squared relationship. A large ratio is mathematically possible. We have added the exact identity in Appendix H.3, Eq. (44): for the same residual collection with positive MAE, MSE divided by squared MAE equals one plus the variance of absolute residuals divided by their squared mean. The variance uses the population-form denominator. Multiplying every residual by the same nonzero constant leaves this ratio unchanged. Therefore, the ratio alone establishes neither an evaluation error nor a normalization-based explanation.

We have added per-variable MAE/MSE, absolute-error quantiles, and the largest 1% residuals' contribution to total squared error in Appendix H.3. These diagnostics connect the ratio to the actual residual distribution and complement the evaluator audit.

3. Published-baseline comparability and repeated entries. Appendix B.2 states that baselines were retrained using official implementations and recommended configurations. We have added a matched-protocol iTransformer reproduction in Table 15(b), documenting splits, input length, horizons, target selection, normalization, evaluation scale, and aggregation. A published single-horizon value and a four-horizon average are not interchangeable.

We have added an entry-level provenance audit in Appendix H.2 and Table 15(c), linking questioned results to source runs, checkpoints, prediction files, corruption realizations, and full-precision metrics. It explicitly covers DeepEDM/BiTGraph on ETTh1 in Table 5, DeepEDM/PDF on ETTh1 in Table 6, and xPatch on ETTh2 across Tables 5 and 6.

We have also included the cross-table discrepancy for Weather with Gaussian noise at a prediction horizon of 192: Table 2 prints MSE 0.952, whereas Table 8 prints 1.052. The audit records the source-derived value and the corresponding table updates. Repeated values are assessed against source records; rounding is not used as a blanket explanation.

We have aligned the affected tables and quantitative statements with the audited records and documented the consequences for comparisons in Appendix H.2. The abstract's original 20.7% refers to the five datasets with reported MSE improvements; the revised presentation explicitly identifies the aggregation scope.

4. Final no-PDR errors and direct mechanism evidence. Figure 2 provides component-removal trajectories and iAUC. Appendix D.1, Figure 3 additionally reports average ETTh1 MAE/MSE reductions of 1.14%/1.12% at a diffusion-loss weight of 0.005 relative to zero weight.

We have added final test MAE/MSE under missingness and noise in Section 5.4 and Table 14(a) for full FenDiM, identical-augmentation/no-PDR, matched deterministic refinement, zero diffusion-loss weight, and reconstruction-trained refinement, with uncertainty and computational costs.

We have separately added paired-reference latent and relational diagnostics in Section 5.6, Appendix G, and Table 13, including the empirical contraction ratio in Eq. (41). Figure 6 visualizes relational changes and corruption-response curves. These distinguish component effectiveness from dependency rectification.

We respectfully request reassessment based on the clarified protocol, numerical audit, and controlled evidence.

Thanks, your response is reasonable for some of my concerns. But dependency rectification is still main weakness and not be address for me. So, i maintain my score.

Reviewer_JraGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FenDiM, a time series forecasting framework motivated by the argument that imperfect observations, such as missing values, noise, and localized distortions, can induce spurious temporal dependencies that degrade forecasting performance. Rather than directly reconstructing corrupted observations, the authors propose prediction-oriented dependency rectification to recover forecasting-relevant representations.

FenDiM combines three main components: (1) Progressive Dependency Rectification (PDR), which applies conditional diffusion with a Mamba-based denoiser in the latent representation space; (2) Frequency-Compensated Temporal Learning (FCTL), which incorporates frequency-domain magnitude and phase information to calibrate temporal representations; and (3) a Mamba-based forecasting backbone. These components are jointly optimized using prediction, diffusion denoising, and spectral consistency losses.

Experiments on six standard time series forecasting benchmarks demonstrate competitive performance under standard observations and synthetically introduced missingness, Gaussian noise, and compound corruptions.

Critical strengths
  1. Strong empirical performance across standard and corrupted forecasting benchmarks.

The proposed method demonstrates competitive forecasting accuracy across six widely used benchmarks, achieving the best horizon-averaged MSE on five of them under standard observation settings. The experiments also cover several levels of missingness and Gaussian noise, providing empirical evidence of robustness to certain forms of input corruption. These results suggest that the proposed architecture is an effective forecasting model, regardless of whether the specific dependency rectification mechanism is responsible for the improvements.

  1. A potentially useful integration of complementary modeling techniques.

The combination of conditional diffusion, frequency-domain conditioning, and Mamba-based temporal modeling is technically interesting. In particular, using frequency-domain information to complement potentially unreliable time-domain representations is a reasonable design choice. The inclusion of Mamba also helps address the computational overhead associated with iterative diffusion-based processing. The ablation and efficiency analyses provide some evidence that these architectural components contribute to forecasting performance and computational practicality.

Critical weaknesses
  1. The central concept of dependency rectification is insufficiently defined and validated.

The paper's primary contribution is framed around the idea that imperfect observations induce spurious predictive dependencies and that the proposed method explicitly rectifies them. However, the key concepts—predictive dependencies, temporal dependencies, dependency bias, and dependency rectification—remain insufficiently operationalized.

Although the introduction provides an intuitive description of predictive dependencies, it is unclear what specific relationships are being rectified, how spurious dependencies are distinguished from forecasting-relevant ones, or how the degree of dependency bias can be measured. Appendix F introduces a generic relational representation, (A(Z)=ϕ(Z)ϕ(Z)⊤A(Z)=\phi(Z)\phi(Z)^\top), but this formulation is not explicitly connected to the implemented rectification objective or empirically evaluated.

More importantly, the proposed PDR objective does not directly establish that corruption-induced dependencies are corrected. The diffusion denoiser is trained to predict artificially injected Gaussian noise from a latent representation (Z_0) that is itself obtained from potentially corrupted observations. Consequently, minimizing the diffusion denoising loss does not necessarily remove the spurious dependencies already embedded in (Z_0). While the joint forecasting objective may encourage more predictive representations, this alone does not demonstrate dependency rectification as a distinct mechanism.

  1. The experimental evaluation does not adequately establish the proposed problem or isolate the claimed mechanism.

The paper argues that observation-induced dependency bias is a significant limitation of existing forecasting methods. However, the experiments rely on conventional time series forecasting benchmarks without establishing whether these datasets inherently contain the types of observation distortions or spurious dependencies motivating the study.

Although the authors evaluate robustness by synthetically introducing random missingness and Gaussian noise, these experiments primarily demonstrate forecasting performance under input corruption. They do not directly establish that the corruption induces the hypothesized dependency bias, or that FenDiM improves forecasting by correcting such dependencies. Furthermore, localized distortions, which are explicitly emphasized in the motivation, are not independently evaluated.

This distinction is important because robustness to corrupted inputs does not necessarily imply successful dependency rectification. A model may achieve improved robustness through regularization, better feature extraction, or greater representational capacity without explicitly identifying or correcting spurious temporal relationships.

The ablation studies demonstrate the utility of the proposed components, but they do not isolate the claimed rectification mechanism from these alternative explanations. In particular, the paper does not quantify dependency distortion before and after PDR or establish that improvements in dependency quality are responsible for the forecasting gains.

As a result, the experiments support the effectiveness of FenDiM as a forecasting architecture, but do not sufficiently substantiate the paper's central claim that prediction-oriented dependency rectification is the underlying reason for its superior performance.

Questions for authors
  1. [Clarification] How are predictive dependencies, dependency bias, and dependency rectification operationally defined in the proposed framework?

The paper uses these concepts as its primary motivation, but their precise meanings and distinctions remain unclear. In particular, how does the proposed method distinguish spurious dependencies from forecasting-relevant ones? Is the relational representation (A(Z)) introduced in Appendix F instantiated in the actual model, or is it intended solely as an analytical abstraction? Clarifying what specifically constitutes a successfully rectified dependency is essential for assessing the claimed contribution.

  1. [Mechanistic validation] What evidence establishes that PDR actually corrects corruption-induced dependencies rather than performing generic latent representation regularization?

Since the diffusion denoising objective predicts noise artificially added to an already potentially corrupted latent representation, it is unclear why this process should specifically eliminate observation-induced spurious dependencies. Can the authors provide direct evidence that the rectified representation is closer to the clean dependency structure than the original representation? Relatedly, is the contraction assumption in Eq. 30 empirically supported by the proposed model?

  1. [Experimental validation] How do the authors establish that the observed forecasting improvements are attributable to the proposed dependency rectification mechanism?

Do the original benchmark datasets exhibit identifiable observation-induced dependency bias, and is there evidence that such bias adversely affects the compared forecasting models? For the synthetic corruption experiments, can the authors demonstrate that dependency distortion actually occurs and is subsequently reduced by FenDiM? In particular, what evidence distinguishes the benefits of targeted dependency rectification from those of conventional denoising, data augmentation, or additional representation learning?

2 comments
AuthorsOpenReview ↗

Clarifying Dependency Rectification: Operational Definition and Direct Mechanistic Validation

Thank you for your careful review and constructive comments. We have revised the definitions, theoretical interpretation, and mechanism-specific evaluation as detailed below.

1. Operational meaning of dependency rectification. We have clarified the definition in Section 3, after Eq. (2), and Section 4.2, after Eq. (8). The predictive relations introduced in Section 1, lines 46-50, are encoded in the initial latent representation in Eq. (3) and refined through Eq. (8). FenDiM does not explicitly estimate a causal graph. The relational matrix in Appendix F, Eq. (25) represents latent relations for analysis; it is not an additional optimization target.

We have defined rectification operationally as reducing corruption-induced relational deviation while preserving forecasting usefulness. Relational similarity and forecasting accuracy are evaluated separately, so neither is presented as sufficient evidence of causal-structure recovery.

2. Diffusion supervision and forecasting guidance. We have revised Sections 4.2 and 4.4 to distinguish observation corruption already encoded in the historical representation from the artificial Gaussian noise predicted by Eq. (17). The complete objective in Eq. (19) also includes clean-future prediction and spectral losses, Eqs. (16) and (18). Thus, the method is trained for forecasting; the auxiliary diffusion loss alone is not claimed to remove observational corruption.

We have expanded the training/sampling description to identify which components receive gradients from each loss, including any detached operations, and to specify the two-step sampler's initialization, timesteps, and updates.

Existing evidence includes component removal in Figure 2 and the zero-diffusion-weight comparison in Appendix D.1, Figure 3. On ETTh1, a diffusion-loss weight of 0.005 reduces average MAE/MSE by 1.14%/1.12% relative to zero weight. This supports the auxiliary objective's contribution in that experiment.

3. Direct tests and alternative explanations. We have added paired-input diagnostics in Section 5.6 and Appendix G.1, reported in Table 13 and Figure 6. With model weights fixed, each historical window is processed before and after added corruption. The encoder output before added corruption serves as the reference. A fixed, nonlearned map uses temporally centered, row-normalized tokens to construct the relational matrix. Eqs. (39)-(40) define this map and the off-diagonal relational distance, which is measured before and after refinement.

Eq. (41) defines an empirical contraction diagnostic: the summed squared Frobenius distances between rectified representations and their paired reference representations, divided by the corresponding sum before rectification. A ratio below one indicates aggregate latent-error reduction for the evaluated pairs. Eq. (42) separately compares rectified reference and corrupted inputs to measure same-stage relational sensitivity. Paired sampling uses matched random draws; near-zero denominators are flagged. Figure 6 presents relational maps and aggregate corruption-response curves, alongside forecasting errors.

We have added controlled final-test comparisons in Section 5.4, Appendix G.2, and Table 14(a): full FenDiM, no PDR with identical augmentation, a capacity-matched deterministic refiner, PDR with zero diffusion-loss weight, and reconstruction-trained refinement. The table reports MAE/MSE, uncertainty, parameters, and inference costs. The appendix specifies supervision, freezing, and training budgets. These controls test augmentation, capacity, the diffusion objective, and prediction-oriented training separately.

4. Theoretical scope and localized corruption. We have clarified in Appendix F, Eqs. (30)-(31) that contraction is a sufficient-condition assumption and the risk bound is conditional on it. Table 13 provides an empirical diagnostic; the theorem itself is not empirical evidence. We have also made the mask dependence explicit when comparing encoder outputs from corrupted, masked histories and their unmasked reference histories.

We have clarified in Section 3 and Appendix G that the reference is the recording before added corruption, without assuming physically noise-free measurements. Beyond the random corruption in Tables 5-9, we have added contiguous missing blocks and localized offsets in Appendix G.3 and Tables 13-14, specifying interval length, affected variables, perturbation magnitude, and randomization, while keeping future targets unchanged.

We respectfully request reassessment in light of these clarified claims and the added mechanism-specific and controlled evaluations.

Thank you for the detailed response and additional experiments. The revisions address several of my concerns and strengthen the empirical validation. However, it remains unclear whether the proposed relational similarity metric adequately captures predictive dependencies and whether the performance gains can be specifically attributed to dependency rectification. Therefore, I will maintain my score.

Comments on the paper

1 comment
Area_Chair_SD5WOpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.