acceptodds
Withdrawn from ICLR 2027

DIAMOND: Diffusion-Informed Patch Masking with Counterfactual Auditing for Time-Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Masking can improve time-series forecasting by reducing reliance on redundant historical information, but a fixed masking ratio may not suit every series, and masking prediction-relevant patches can harm accuracy. We propose DIAMOND, a training-time framework that adapts both the masking ratio and patch selection to each input series. A conditional diffusion branch estimates how readily historical patches can be reconstructed from visible context and provides evidence for a sample-specific masking budget. A scalar predictor estimates the masking ratio, while a direction predictor ranks candidate patches. To align patch selection with the forecasting task, DIAMOND periodically restores individual masked patches and measures the resulting change in prediction loss. This counterfactual feedback supervises patch ranking and helps protect patches that are important for forecasting. At inference, the trained forecaster uses the complete historical input without the masking controller or diffusion branch. Experiments across benchmark datasets and forecasting backbones demonstrate consistent improvements over existing approaches. Source code is available at https://anonymous.4open.science/r/Diamond-code-A812.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Adaptive Masking, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5f9mOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DIAMOND, an adaptive patch masking framework for time-series forecasting that combines diffusion-based reconstruction with counterfactual auditing to determine how much and which historical patches to mask. Experiments across multiple datasets and six forecasting backbones show consistent improvements. However, the novelty of the proposed combination and its advantages over simpler adaptive masking strategies remain insufficiently established.

Critical strengths
  1. Clear problem formulation. The paper distinguishes patch recoverability from forecasting relevance, motivating separate decisions for masking ratios and patch selection.

  2. Broad empirical evaluation. Experiments across multiple datasets and six forecasting backbones show consistent improvements over baseline models, supported by ablation studies.

Critical weaknesses
  1. Limited justification of design choices. Although the ablation study supports the overall framework, comparisons with simpler alternatives, such as conventional reconstruction models or a unified masking controller, are missing. The necessity of the specific design choices therefore remains unclear.

  2. Limitations of counterfactual auditing. Restoring patches individually measures their effects under a fixed masking context, potentially overlooking interactions between patches. The reliability of the resulting rankings is not sufficiently investigated.

  3. Limited theoretical support and inconsistent documentation. The theoretical analysis establishes properties of diffusion reconstruction error and EMA smoothing, but does not justify the effectiveness of the overall masking strategy. Additionally, Appendix G.1 contains spectral-scoring terminology inconsistent with the proposed method, raising concerns about the precision of the experimental documentation.

Questions for authors

See the weaknesses.

Reviewer_kbWhOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DIAMOND, a training-time framework that adaptively determines both the masking ratio and patch selection for each input series. A conditional diffusion branch estimates how readily historical patches can be reconstructed from the visible context, informing a sample-specific masking budget. DIAMOND also periodically restores individual masked patches and measures the resulting change in prediction loss. This counterfactual feedback guides patch ranking, helping preserve patches that are important for forecasting. Experimental results demonstrate the framework’s effectiveness.

Critical strengths

The paper is clearly written and easy to follow, with a well-organized structure and a comprehensive experimental evaluation.

Critical weaknesses

Time series forecasting is outside my primary area of expertise. I have contacted the AC to request reassignment to a paper more closely aligned with my expertise but have not yet received a response. Although I have read the paper carefully and found the presentation clear, I do not feel sufficiently qualified to identify substantive technical weaknesses with confidence. I therefore refrain from raising concerns that I cannot adequately substantiate.

I will carefully consider the other reviewers’ assessments during the discussion and encourage the authors to focus their response on the substantive questions and concerns those reviewers raise.

Questions for authors

N/A

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

DIAMOND proposes a training-time framework for adaptive patch masking in time-series forecasting. A conditional diffusion branch estimates patch recoverability, while budget and direction predictors determine how many patches to mask and which to select. Counterfactual restoration assesses whether masking particular patches harms the downstream forecasting objective and feeds this information back to the masking policy. The auxiliary components are used only during training and removed at inference. The evaluation spans forecasting architectures and datasets, with additional component, transfer, and efficiency analyses.

Critical strengths
  1. A well-motivated extension of reconstruction-driven masking. A patch that is easy to reconstruct may still matter for forecasting. The downstream counterfactual loss gives the masking policy a forecasting-specific signal, providing a clear rationale for the proposed design. Restricting the auxiliary modules to training also offers a practical deployment advantage.

  2. Broad empirical coverage with unfavorable cases reported. The evaluation includes six forecasting backbones and nine long-term forecasting datasets, alongside component ablations, additional forecasting settings, and computational analysis. The main backbone–dataset averages are favorable, while the detailed tables also report degradations at individual forecast horizons. These results support the potential utility of the training scheme, although the contribution of each module remains to be established more clearly.

Critical weaknesses
  1. The contribution of counterfactual auditing and the reliability of its gains remain unclear. The direction-ranking loss, protective gate, and budget controller are coupled, but the ablations do not clearly distinguish their contributions, particularly in the ILI example. Some appendix parameters are not explicitly linked to the main optimization objective. The appendix also contains inconsistent summary values, including the iTransformer–Electricity average. All 54 backbone–dataset averages improve over the unmasked backbones, yet 30 of 216 individual-horizon MSE comparisons show degradation. Gains over the stronger masking baseline are also modest in several settings. Results across multiple random seeds in a few representative settings would clarify whether the gains reliably justify the added complexity.

  2. The stated audit and conditioning mechanisms need to be reconciled with the supplied implementation. Equation (13) describes audit pairs within a sample, whereas the supplied code appears to aggregate restoration effects by patch position across a minibatch. This implementation can form ranking pairs even when each sample masks only one patch. Single-patch masking therefore does not, by itself, establish that the direction loss is inactive. In addition, the generic non-PatchTST conditioning path projects the original patches rather than using representations derived solely from the visible context after masking. These differences matter for understanding what the diffusion branch learns and which shared parameters it regularizes.

Questions for authors
  1. Could the authors specify the masking objective and ablation definitions used in Tables 1–3 and 5? Does w/o CF remove direction ranking, the protective gate, or both? Please explain how the additional budget parameters enter training and identify which appendix summaries correspond to the main experiments. A short mapping between configurations and tables would be sufficient.

  2. Could the authors specify the conditioning representations, audit-pair construction, and actual patch and mask counts for the main backbones, particularly ILI+iTransformer? Please explain how the sample-level formulation relates to the batch-level audit procedure in the supplied code. Repeating the central ablation across a few random seeds in representative settings with small gains would help assess robustness.

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

Related papers

Discussion (0)

Sign in to comment.