SHIFT-RESILIENT DIFFUSIVE IMPUTATION FOR VARIABLE SUBSET FORECASTING

1. Several key claims require clarification. In section 4.2, why the divide-conquer strategy can handle the problem of inter-series shift? In section 5, why did the author choose the meta-learning strategy to mitigate the intra-series shift? The authors should explicitly compare and contrast the roles of Equations 11 and 13, and explain why both are necessary.

2. Two-stage training process may lead to error accumulation.

3. The presentation of core functional modules is mainly based on text description accompanied by snippets of equations that only show the intermediate transformation steps whereas the whole picture is lost. For example, it is not clear how Equation 11 and Equation 5 are constructed after reading the methodology section. The authors are supposed to provide a precise mathematical description of the entire data transformation process, i.e., how the inputs are transformed into the output step by step.

1. Writing Issues:

1.1. Structural Inconsistency: While the paper claims to "categorize and provide a comprehensive analysis of two distinct types of shift in VSF tasks," this fundamental analysis is insufficiently developed in the main text. Such a core concept warrants a dedicated section rather than being merely mentioned in the contributions or inferred from conclusions.

1.2. Insufficient Technical Details: The paper omits crucial specifications regarding model parameters, training complexity, and inference costs. Given the framework's apparent computational intensity, these quantitative metrics are essential for evaluating its practical applicability.

1.3. Results Presentation: The experimental results should prioritize comparisons with other imputation methods and alternative VSF solutions, rather than focusing on various forecasting model combinations, given the two-stage nature of the proposed approach.

2. Experimental Concerns:

2.1. Counter-intuitive Results: The performance reported in Table 1, where the model surpasses the Oracle (full data) baseline with 85% missing data, raises significant concerns about the experimental setup or evaluation metrics. This counter-intuitive outcome requires thorough explanation or methodological reexamination.

2.2. Fairness in Evaluation: The meta-learning adaptation during testing introduces a fundamental fairness issue. The proposed method uniquely accesses and adapts to test data, while all other baselines (including Oracle) operate without such access. This creates an incomparable evaluation framework, making it impossible to fairly assess the model's actual contribution to VSF.

2.3. Limited Scope of Comparison: The paper focuses exclusively on imputation-based methods, overlooking significant non-imputation solutions (e.g., Chauhan 2022, GinAR from KDD2024). This limitation is particularly problematic given the fundamental differences between traditional imputation and VSF scenarios.

2.4. Incomplete Baseline Selection: The omission of recent relevant work (e.g., PriSTI 2023(mentioned in section 4.1)) from the comparative analysis compromises the evaluation's comprehensiveness.

Weaknesses

1. Technical Workflow: The reviewer questions the suitability of directly applying diffusion models for missing data imputation. As noted in reference [1], "the inherent mismatch in learning objectives between generation and imputation can impact imputation performance in diffusion models. Diversity, which is essential in generation tasks, conflicts with the imputation task's need for accuracy rather than variation." Since diffusion models are sensitive to initial noise $x_T$, which facilitates diverse sample generation, the reviewer questions whether diffusion models are inherently suitable for precise imputation tasks.
2. Divide-and-Conquer Strategy: The “divide-and-conquer” strategy in Figure 1 appears to resemble the residual modules in the N-BEATS model [2]. How does this strategy differ in innovation from similar approaches such as N-BEATS?
3. Missing Data Patterns: Missing data patterns are typically categorized as missing at random, missing completely at random, and missing not at random [3]. Is the distribution shift primarily related to the “missing not at random” pattern? If so, it would seem crucial to incorporate the missing data mechanism during the modeling stage.
4. Figure Reference: Figure 1 is not referenced in the manuscript text.
5. Experimental Design: The sensitivity of the weight parameter $\overline{\omega}$ is not evaluated with respect to various strengths. Adding a sensitivity analysis for this parameter would enhance the experimental design.
6. Experimental Presentation: Error bars are absent in Figures 4 and 5, which would help demonstrate error bars in the experimental results.
7. Typographical Errors: For example, in the problem formulation, should $\mathcal{F}\_\theta$ be corrected to $\mathcal{F}\_\Theta$?

References:
[1] Self-Supervision Improves Diffusion Models for Tabular Data Imputation. CIKM 2024, Long Paper Track
[2] N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. ICLR 2020, Main Track
[3] Inference and missing data. Biometrika 1976

1.RELATED WORK is too simple. Many related studies are missing. The significant difference between the related studies and the proposed method of this paper is not analyzed, making the novelty of this paper be not clear. 2. The time complexity of the proposed algorighm is not analyzed. 3. The details of the adopted datasets are missing, such as the number of features and instances, the time interval between two instances. 4. The latest comparison model is MSTGCN Jia et al. (2021), more recent models are desired to be compared. Besides, some imputation methods could be compared, such as the methods introduced in Section 2.1 and “Missing Value Imputation for Mixed Data via Gaussian Copula, in KDD 2020”, “Transformed Distribution Matching for Missing Value Imputation, in ICML 2023”. 5. What is the impact of the hyperparameter ϖ. 6. Presentation issues. 1） line 124, page 3: In 2020, Lin et al. (2023) presented…, is it in 2023? 2) Figure 1 is not cited in the main text.