acceptodds
Under review as a conference paper at ICLR 2027

Domain Generalization under Sampling Pattern Shifts in Irregular Time Series

Abstract

Irregularly sampled multivariate time series (ISMTS) are prevalent in real-world applications, where both observation times and available measurements can vary substantially across domains. While recent models increasingly exploit such sampling information for prediction, its robustness under sampling pattern shifts remains underexplored. We introduce HAR-C, to the best of our knowledge the first controlled benchmark for sampling pattern shifts in ISMTS, and show that sampling shifts alone can substantially degrade performance, induce sampling-specific shortcuts, and remain challenging for existing domain generalization (DG) methods. Motivated by these findings, we propose PRISM, a DG framework that first learns complementary feature-centric and sampling-centric representations without task labels, and subsequently performs robust supervised training across diverse sampling variations to discourage brittle shortcut reliance. Extensive experiments on controlled and real-world ISMTS benchmarks demonstrate that PRISM consistently improves robustness to unseen sampling shifts over existing methods. Our code is available at https://anonymous.4open.science/r/PRISM.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series, Sampling Pattern Shifts, Domain Generalization

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_2NXhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies domain generalization classification under sampling pattern shifts in irregularly sampled time series. The authors focus on modern irregular time series models. First, the authors developed a benchmark dedicated to sampling pattern shifts in Irregularly sampled multivariate time series (ISMTS). The benchmark simulates different designed sampling shifts and it was used to test cross shifting patterns and domains. Based on the findings the authors proposed a new method PRISM, that aims to address the limitations of existing work.

The key contributions are the proposed benchmark, the analysis, and the proposed method.

Critical strengths

Pros:

  1. The analysis is interesting. Some insights (e.g., Fraction of shifted test samples and Shortcut strength analysis in figure 2) can be valuable for the research community.
  2. The proposed method shows good performance and also under ablation study. Table 1 shows the proposed method is quite strong under the synthetic environment. Tables 2 and 3 also shows the effectiveness of the method.
Critical weaknesses

Cons: The analysis seems to be less supported by the claims and for the model design. Below are some specific feedbacks.

  1. For the analysis:
    1. The analysis (sec. 2) in general is a bit hard to understand before the introduction of the task in Sec. 3.1.
    2. The section of ‘Real-world sampling processes are predictive, yet domain dependent’ can be improved.
      1. In Figure 4, the authors compare results of ‘observation masks alone’ and ‘full input’. It would be helpful to explain what is the full input. The claim that ‘mask-only prediction markedly under cross-region transfer’ also seems less supported.
      2. For line 185 “sampling patterns can be both predictive and domain dependent—the two ingredients that make sampling-specific shortcuts brittle under domain shift”, the authors use the analysis of Figure 4 and 1 to explain the results in figure 3. The connections are not clear and no evidence directly this.
      3. ‘Label’ is used throughout in sec. 2.2 but there is no definition. According to sec. 3.1, does label mean m and t which encode data irregularity? Or is the class to be predicted?
  2. For the model design:
    1. The motivation of sec. 3.2 Complementary representation pretraining can be improved. It is not very clear how the controlled representation learning can help mitigate the shortcut and improve DG. How the learn representation are invariant to sampling patterns (line 238).
  3. Other things that can be clearer
    1. The settings of ‘bias-conflicting accuracy’ and the ‘mask-following rate’ for Fig. 4.
    2. Eq. 1, vl,d\mathbf{v}_{l,d} means the ground truth features. Thus, it would be better to be more specific than stating “by randomly hiding additional observations.” (In line 247)
      1. Explain what the ‘reference’ mean in “The Regular process serves as a reference condition” In line 148.
    3. It would also be more specific for the domain generalization task in the introduction, cause there are also other tasks than the classification studied in the work, e.g., for time series forecasting, e.g., generalization across domains or over time.

The anonymized code repository also seems expired.

Questions for authors

Please see Critical Weaknesses*

5 comments

We sincerely thank the reviewer for the thoughtful, detailed feedback; addressing the comments on the writing has substantially improved the clarity and quality of our paper. We especially appreciate the positive evaluation of (1) the interesting analysis, with insights such as the fraction of shifted test samples and the shortcut strength analysis (Figure 3) being valuable to the community, and (2) the strong performance of our method, including in the ablation study (Tables 1–3). We address all concerns below.


[W1]

The analysis (sec. 2) in general is a bit hard to understand before the introduction of the task in Sec. 3.1.

‘Label’ is used throughout in sec. 2.2 but there is no definition. According to sec. 3.1, does label mean m and t which encode data irregularity? Or is the class to be predicted?

We thank the reviewer for these suggestions. We have moved the problem setup (formerly Section 3.1) to the start of Section 2, so the task and all symbols, now used in Section 2, are defined before the analysis (former Sections 3.2 and 3.3 are now 3.1 and 3.2). "Label" denotes the class label y of the downstream task (e.g., in-hospital mortality on eICU, the activity on HAR-C), never the sampling information m or t, as the problem setup now states.


[W2]

The section of ‘Real-world sampling processes are predictive, yet domain dependent’ can be improved. In Figure 4, the authors compare results of ‘observation masks alone’ and ‘full input’. It would be helpful to explain what is the full input. The claim that ‘mask-only prediction markedly under cross-region transfer’ also seems less supported.

Thank you for these suggestions.

Full input is the model input used throughout the paper: observed values, observation mask and timestamps. In "mask only", the observed values are replaced by the observation mask, so the model sees where and when, but not what, was measured. We have added this definition to Section 2.2.

Predictive, yet unstable across regions. With 50% AUROC as random guessing, observation masks are clearly predictive, exceeding chance by 15.3±0.3 points in-domain on average, but lose 5.8±0.5 points in the worst other region (Figure 4, left), about a third of this margin. On the same test set, too, a model trained in another region is worse than the in-region model in 11 of 12 region pairs (AUROC, 5 seeds; columns are target regions):

Target regionNortheastWestSouthMidwest
Mask only, trained in region72.6±1.263.8±1.260.7±1.364.3±1.5
Mask only, trained elsewhere68.9±0.862.0±0.559.9±0.260.0±0.5
Full input, trained in region78.8±2.070.1±1.469.7±0.772.3±1.5
Full input, trained elsewhere74.5±1.467.4±0.768.1±0.469.2±0.4

We agree the original text did not quantify this drop and have added these numbers to Section 2.2.


[W3]

For line 185 “sampling patterns can be both predictive and domain dependent—the two ingredients that make sampling-specific shortcuts brittle under domain shift”, the authors use the analysis of Figure 4 and 1 to explain the results in figure 3. The connections are not clear and no evidence directly this.

Thank you for pointing this out. We understand this reading: the eICU paragraph announced "the same phenomenon" and called its findings "the two ingredients that make sampling-specific shortcuts brittle", as if they explained Figure 3. However, Figure 3 is a self-contained controlled HAR-C experiment that directly shows this brittleness; Figures 1 and 4 (left) only check whether its conditions also arise in real data. Our logical flow is:

  1. Sampling varies across real-world domains (Figure 1), but many factors change together there, so we isolate sampling pattern shifts in our HAR-C.
  2. There, sampling pattern shifts alone degrade performance, and a label-associated pattern becomes a shortcut that fails once the association changes (Figure 3; ERM: 94.0±2.4 bias-aligned vs. 42.6±1.9 bias-conflicting accuracy, Table A6).
  3. On eICU, observation masks alone are predictive (15.3±0.3 AUROC points above chance) but lose 5.8±0.5 points in the worst other region (Figure 4, left; W2), so such a shortcut would be unstable.
  4. Existing DG methods do not resolve this shortcut (Figure 4, middle and right).

We have revised line 185 and the opening of the eICU paragraph accordingly.

[W4]

The motivation of sec. 3.2 Complementary representation pretraining can be improved. It is not very clear how the controlled representation learning can help mitigate the shortcut and improve DG. How the learn representation are invariant to sampling patterns (line 238).

Thank you for this question. First, PRISM does not aim for invariance to sampling patterns: as stated in line 238, invariance "may discard useful predictive cues", such as the more frequent monitoring of deteriorating patients in eICU (Section 2.2). PRISM's design rationale is:

  1. Pretraining without task labels. Under task supervision, sampling patterns become shortcuts as simple, directly observable cues (ERM: 94.0±2.4 bias-aligned vs. 42.6±1.9 bias-conflicting accuracy, Table A6), which motivates pretraining rich representations without task labels.
  2. Separate encoders. Both kinds of information are needed and complementary: E_feat recovers removed feature values (Eq. 1), E_samp reconstructs the observation mask (Eq. 2); a larger single encoder trained with both objectives underperforms PRISM on HAR-C and eICU (7rtk, W1).
  3. Worst-view objective downstream. Even with separate encoders, the classifier may ignore the feature representation and use E_samp alone, so PRISM minimizes the worst loss over K sampling variations (Section 3.2), making reliance on a cue that some variations disrupt costly (HAR-C: 72.02±0.43 without variations vs. 80.57±0.19, Table 3).

This design directly reduces shortcut reliance: on the shortcut benchmark (ρ=0.9) with the same sampling variations, complementary pretraining raises bias-conflicting accuracy from 38.30±1.30 to 49.11±0.74 and lowers the mask-following rate from 0.498±0.011 to 0.405±0.015 (p-value=1e-5 and 1e-4; 7rtk, W2), while a single encoder with the same variations follows the sampling cue even more (0.556±0.016). Table 3 shows the DG gain from complementary pretraining. We have revised the start of Section 3.1 accordingly.


[W5]

The settings of ‘bias-conflicting accuracy’ and the ‘mask-following rate’ for Fig. 4.

Thank you for pointing this out. We have added the following definitions to Section 2.2, where Figure 3 (varying ρ) and Figure 4 (ρ=0.9) are discussed.

  • Setting. Each of the six HAR classes is tied to a distinct sampling pattern, all with the same observation budget; each training sample is observed with its class's pattern with probability ρ (ρ=0.9 in all four source environments for Figure 4) and with one of the other five otherwise.
  • Test sets. The same test samples are observed with their class's pattern (bias-aligned), a pattern drawn independently of the label (unbiased), or the pattern of a class y_m other than the true label y (bias-conflicting).
  • Metrics. Bias-conflicting accuracy is the accuracy on the bias-conflicting set; the mask-following rate is the fraction of bias-conflicting samples predicted as y_m, i.e., following the sampling cue rather than the true label (chance level: 1/6). Less shortcut reliance means higher bias-conflicting accuracy and a lower mask-following rate.

[W6]

Eq. 1, [...] means the ground truth features. Thus, it would be better to be more specific than stating “by randomly hiding additional observations.” (In line 247)

Thank you for the suggestion. In Eq. 1, v denotes the ground-truth feature values, and the loss covers only entries observed in x but removed by T (m = 1, m̃ = 0). Were entries visible in x̃ included, E_feat could minimize the loss by an identity mapping that copies its input, capturing little beyond raw values; recovering removed values instead requires encoding how values evolve over time and relate across features, yielding a richer representation. By "randomly hiding additional observations" (line 247), we meant hiding some observed entries of x (m = 1) beyond those already unobserved (m = 0); as this was unclear, the revised sentence states that T removes the observations of x at randomly selected time points and features, with the drop ratios at the end of Section 3.1 (Table A5).


[W7]

Explain what the ‘reference’ mean in “The Regular process serves as a reference condition” In line 148.

By "reference", we meant that performance is maintained under regular sampling, unlike under the other shifts (ERM: 80.97±0.81% vs. 79.92±0.55% in the source domain (Random), Table 1; 54–69% under the other shifts). Regular sampling observes all features at evenly spaced time points, differing from Random only in this spacing: in both, all features are observed together and observations span the whole sequence. As the term was unclear, we have replaced it with this explanation (Sections 2.1 and 2.2).


[W8]

It would also be more specific for the domain generalization task in the introduction, cause there are also other tasks than the classification studied in the work, e.g., for time series forecasting, e.g., generalization across domains or over time.

To make our scope explicit, we have revised the introduction (Section 1) and the problem setup (now at the beginning of Section 2) to state that we study domain generalization for classification of irregularly sampled multivariate time series.


[W9]

The anonymized code repository also seems expired.

Thank you for pointing this out. The anonymous repository has been restored with the code for all new experiments (https://anonymous.4open.science/r/PRISM-8C66), and the link is updated in the OpenReview abstract and the revised paper (Abstract, Section 4.1, Reproducibility Statement).

Reviewer_6Qi7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper first evaluates the impact of irregular sampling in time series by providing an empirical analysis on the well-known HAR dataset. From this assessment, the paper introduces a novel method denoted PRISM for learning robust representations under sampling pattern shifts by combining a feature-centric and a sampling-centric representation.

Critical strengths

The paper seems to be well written. The paper first provides an empirical analysis and then introduces a novel method to address irregular sampling.

Critical weaknesses

The first part of the paper studies empirically the influence of irregular sampling by considering the HAR data. The authors fail to cite related work, such as the following work that also considers the HAR to evaluate the influence of irregular sampling: Liu, M., Geißler, D., Bian, S., Zhou, B., & Lukowicz, P. (2025). Assessing the Impact of Sampling Irregularity in Time Series Data: Human Activity Recognition As A Case Study. In 2025 IEEE International Conference on Pervasive Computing and Communications Workshops and Other Affiliated Events (PerCom Workshops) (pp. 311–316). IEEE.

The analysis of sampling pattern shifts only considers the HAR dataset. It is not clear if the obtained empirical results would generalize to other datasets. Moreover, some of the obtained results on irregular sampling seem evident. Having other datasets would confirm or contradict the resulting discussion.

Related work is relatively relevant. However, while several related papers are from 2023 and before, there is only one paper from 2024, one from 2025, and another one from 2026. Moreover, these recent papers are not much related to this work (e.g. the use of a language model). At the end, the paper is missing many recent related works from the more recent literature.

The experimental analysis is also missing comparative analysis with the state of the art. The most recent compared methods are Diversify (2023) and ManyDG (2023), which are a bit old.

In many experiments, the proposed method PRISM performs nearly identically to MLDG. It is not clear why both share roughly similar performance, considering that MLDG explores a meta-learning strategy. It is also remarkable that MLDG is pretty old, from 2018.

Questions for authors

none

7 comments

We sincerely thank the reviewer for the constructive feedback and for noting that the paper is well written and first provides an empirical analysis before introducing a novel method for irregular sampling. We address all concerns below.


[W1]

The first part of the paper studies empirically the influence of irregular sampling by considering the HAR data. The authors fail to cite related work, such as the following work that also considers the HAR to evaluate the influence of irregular sampling: Liu, M., Geißler, D., Bian, S., Zhou, B., & Lukowicz, P. (2025). Assessing the Impact of Sampling Irregularity in Time Series Data: Human Activity Recognition As A Case Study. [...]

We thank the reviewer for this valuable literature, now cited and discussed in Section 2.1 and Appendix B. We acknowledge that [1] studied the effect of sampling irregularity on HAR before our work.

[1] is a preliminary workshop study of two simple irregularity types, (i) random timestamp offsets, similar to jittering, which injects random noise into the values [2], and (ii) randomly dropped observations [3, 4], extending perturbations common on other time series datasets to HAR. Random dropping also reduces the available information, so its effect cannot be separated from that of the sampling pattern change.

In contrast, to our knowledge, HAR-C is the first benchmark that systematically isolates sampling pattern shifts. In HAR-C:

  1. Controlled observation budget. All patterns keep the same 20% observation budget (Section 2.1), separating sampling pattern shifts from a reduced budget; in the matched-pattern control (7rtk, W3), training on the test pattern recovers most of the degradation (12.4±0.6 of 14.0±0.7 points), which thus stems from the sampling pattern shift itself.
  2. Diverse sampling pattern shifts. 7 target shifts at varying severity (Figure 2): regular sampling (Regular), desynchronization across features (Desync), heterogeneous feature-wise sampling (Fixed-Feat. and Rand-Feat.), and temporal concentration toward different parts of the sequence (First, Last, and Mid).
  3. Label-correlated sampling. As models are sensitive to these shifts, we further tie each class to a distinct sampling pattern; models increasingly follow the sampling-associated label and fail when the association breaks (Figure 3; ERM: 94.0±2.4 bias-aligned vs. 42.6±1.9 bias-conflicting accuracy, Table A6).

[1] Liu et al. "Assessing the Impact of Sampling Irregularity in Time Series Data: Human Activity Recognition As A Case Study." PerFail Workshop at PerCom 2025.
[2] Um et al. "Data Augmentation of Wearable Sensor Data for Parkinson's Disease Monitoring using Convolutional Neural Networks." ICMI 2017.
[3] Kidger et al. "Neural Controlled Differential Equations for Irregular Time Series." NeurIPS 2020.
[4] Oh et al. "Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data." ICLR 2024.

[W2]

The analysis of sampling pattern shifts only considers the HAR dataset. It is not clear if the obtained empirical results would generalize to other datasets. Moreover, some of the obtained results on irregular sampling seem evident. Having other datasets would confirm or contradict the resulting discussion.

We thank the reviewer for this suggestion. Of the four Section 2.2 analyses, the real-world one already uses eICU (Figure 4, left); we repeated the other three, controlled analyses on PAM with the same protocol and observation budget, and all HAR-C findings hold (Table A15), including that no DG baseline of Figure 4 improves bias-conflicting accuracy significantly over ERM (all p-value≥0.11).

1. Sampling pattern shifts alone degrade prediction. ERM trained on Random (82.0±0.2), with an increasing fraction of test samples shifted (as in Figure 3):

Shift Type0.20.40.60.81.0
Regular82.3±0.382.2±0.382.3±0.482.4±0.482.5±0.4
Desync81.5±0.480.6±0.479.8±0.578.8±0.578.0±0.6
Fixed-Feat.76.6±0.870.8±1.465.5±2.160.0±2.754.4±3.5
Rand-Feat.78.1±0.473.9±0.469.6±0.665.3±0.760.9±0.8
First81.4±0.480.6±0.579.7±0.679.1±0.778.3±0.8
Last80.8±0.379.7±0.378.5±0.477.7±0.476.4±0.6
Mid81.4±0.380.8±0.380.2±0.279.9±0.379.3±0.3

As on HAR-C, accuracy drops as more samples are shifted, most under the feature-wise sampling pattern shifts, but is maintained under regular sampling.

2. Sampling patterns become shortcuts when correlated with labels. As on HAR-C, each of six activities is tied to one of the six shortcut-benchmark sampling patterns with probability ρ (the grid of Figure 3), each keeping exactly 50% of the originally observed entries of a window (single-activity windows; falling, with only 23, is excluded):

ρBias-alignedUnbiasedBias-conflictingMask-following rate
1/687.3±0.587.7±0.488.7±0.40.007±0.001
2/690.8±0.387.8±0.388.3±0.30.012±0.001
3/693.7±0.287.4±0.387.3±0.40.014±0.001
4/695.3±0.285.3±0.585.8±0.60.024±0.001
5/697.1±0.281.4±0.482.0±0.80.048±0.005

As ρ increases, bias-aligned accuracy and the mask-following rate rise while bias-conflicting accuracy falls.

On the results seeming evident. Training on the test pattern recovers 12.4±0.6 of the 14.0±0.7-point average HAR-C drop (W1), so most of it is due to the sampling pattern shift itself, not to the observation budget, which is fixed. We have added these results as Table A15 (Appendix D) and to Section 2.2.

[W3]

Related work is relatively relevant. However, while several related papers are from 2023 and before, there is only one paper from 2024, one from 2025, and another one from 2026. Moreover, these recent papers are not much related to this work (e.g. the use of a language model). At the end, the paper is missing many recent related works from the more recent literature.

We thank the reviewer for finding the related work relevant. We have extended it (Appendix B) with recent models for irregularly sampled time series [7–10], which, like the language-model-based models we cited [5, 6], propose or analyze architectures (e.g., patching and graph-based models, adaptive time encodings), and with other recent works, including the DG methods added as baselines (W4).

As noted in Section 1, research on irregularly sampled time series has mainly focused on architectures that directly process irregular time series in their native form, while their robustness under real-world distribution shifts, such as sampling pattern shifts between the training and test domains, has yet to be fully scrutinized. Our work is complementary: PRISM is an architecture-agnostic training framework, not an architecture, and its improvement holds across backbones (Figure 6, Section 4.4). The HAR study suggested by the reviewer [1] is discussed in W1.

[5] Tan et al. "Are Language Models Actually Useful for Time Series Forecasting?" NeurIPS 2024.
[6] Tang et al. "Time Series Forecasting with LLMs: Understanding and Enhancing Model Capabilities." SIGKDD Explorations 2025.
[7] Zhang et al. "Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach." ICML 2024.
[8] Luo et al. "Knowledge-Empowered Dynamic Graph Network for Irregularly Sampled Medical Time Series." NeurIPS 2024.
[9] Luo et al. "Hi-Patch: Hierarchical Patch GNN for Irregular Multivariate Time Series." ICML 2025.
[10] Lee et al. "Adaptive Time Encoding for Irregular Multivariate Time-Series Classification." NeurIPS 2025.


[W4]

The experimental analysis is also missing comparative analysis with the state of the art. The most recent compared methods are Diversify (2023) and ManyDG (2023), which are a bit old.

We agree and added two state-of-the-art DG methods, MEDRO [11] (NeurIPS 2025) and Hierarchical DRO [12] (ICLR 2026), ported from the official code and run under the same protocol as all baselines (Section 4.1; Tables 1, 2 and A7).

HAR-C (acc)eICU (AUROC)PAM leave-fixed-sensors-out (acc)
ERM66.47±0.9579.72±0.1568.11±2.39
GroupDRO66.78±0.6879.02±0.1368.24±2.16
MLDG72.70±0.5480.14±0.0770.70±2.33
MEDRO (NeurIPS'25)69.03±0.8079.66±0.1669.84±1.93
Hierarchical DRO (ICLR'26)70.00±0.7679.18±0.1869.94±2.06
PRISM80.57±0.1982.47±0.0484.10±0.48

Although state-of-the-art under general domain shifts and better than GroupDRO, which they extend (+2.25±0.86 and +3.22±0.98 on HAR-C), both stay below MLDG on every dataset and far below PRISM (all p-value<1e-4), consistent with the other DG baselines (Tables 1 and 2; Section 2.2): robustness to general domain shifts does not carry over to sampling pattern shifts. Both minimize the worst-case risk over source groups (MEDRO with an expert head per environment, Hierarchical DRO with ambiguity sets within groups), guarding against shifts already present among them. A sampling pattern shift, however, brings patterns absent from every source environment (e.g., all HAR-C source environments use Random), and neither method models how observations are sampled. This reaffirms that sampling pattern shifts must be addressed explicitly, as PRISM does.

[11] Jeong et al. "Multi-Expert Distributionally Robust Optimization for Out-of-Distribution Generalization." NeurIPS 2025.
[12] Jo et al. "Mitigating Spurious Correlation via Distributionally Robust Learning with Hierarchical Ambiguity Sets." ICLR 2026.

Reviewer_7rtkOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the domain generalization problem of irregular multivariate time series when sampling patterns change. The authors construct the HAR-C test benchmark, analyzing the model’s reliance on sampling shortcuts by changing sampling positions and using different sampling patterns. Based on this, the authors propose the PRISM method. First, two encoders are pretrained using numerical reconstruction and observation mask reconstruction, respectively, and then the maximum classification loss over multiple randomly sampled observations is optimized. Experiments cover HAR-C, eICU, and PAM, and include component ablations as well as validations on different backbone networks.

Critical strengths
  1. The paper provides a fairly complete problem motivation, and the research question is valuable. Combined with label-related sampling patterns and eICU mask prediction experiments, it allows for a more direct observation of the effects of sampling pattern changes.
  2. The method achieves good experimental results. On HAR-C, PRISM achieves an average accuracy of 80.57%, far higher than the strongest comparative baseline of 72.70%. On eICU, the average AUROC is 82.47%, also outperforming comparative methods. Good results are also observed on PAM and tests with various backbone networks.
Critical weaknesses
  1. Ablation experiments are insufficient. Table 3 shows that adding sampling augmentation on top of the dual-branch pretraining already achieves an average HAR-C accuracy of 79.58%, indicating that sampling augmentation may contribute most of the gain. The current comparisons simultaneously change pretraining, model capacity, augmentation strategy, and optimization objective, so the independent value of dual-branch complementary learning cannot be fully determined.
  2. There is insufficient evidence for PRISM’s mechanism to reduce reliance on sampling shortcuts. In Figure 4, PRISM’s mask-following rate is similar to that of ERM, making it unclear whether the improvement comes from stronger utilization of numerical features or from reduced dependence on sampling shortcuts.
  3. HAR-C controls the number of observations but does not rule out the effects of task-relevant feature changes. Different time periods and sensors may contain varying levels of class-discriminative information; therefore, the same number of observations does not necessarily imply the same prediction difficulty. The performance drop in HAR-C may result from both the failure of sampling shortcuts and the absence of key discriminative information.
Questions for authors
  1. Under the same sampling augmentation, what are the results of using a single encoder with average loss and worst-view loss, respectively? Compared to a dual encoder with a matched parameter count, does complementary pretraining still have a consistent advantage?
  2. Table A8 marks K=1 as 'no sampling change.' According to the formula, K=1 may still apply one random transformation to the input. Does this ablation actually use the original input or a randomly augmented view?
  3. In HAR-C and PAM, how are the source domain labels for each DG method constructed? Are the candidate ranges for hyperparameters, search budget, validation metrics, and model selection procedures consistent?
  4. Analysis of prediction changes when only changing the sampling mode while keeping the base sequence fixed, and comparing it with a single encoder under the same augmentation to more directly test sampling dependence? What model, data split, and reconstruction target were used for the probe in Figure A9?
7 comments

We sincerely thank the reviewer for the precise, constructive questions and the positive evaluation of (1) our complete problem motivation and valuable research question, with the label-related sampling patterns and eICU mask-prediction experiments directly showing the effects of sampling pattern changes, and (2) our strong results across HAR-C, eICU, PAM and various backbones. We address all concerns below.


[W1 & Q1]

Ablation experiments are insufficient. Table 3 shows that adding sampling augmentation on top of the dual-branch pretraining already achieves an average HAR-C accuracy of 79.58%, indicating that sampling augmentation may contribute most of the gain. The current comparisons simultaneously change pretraining, model capacity, augmentation strategy, and optimization objective, so the independent value of dual-branch complementary learning cannot be fully determined.

Under the same sampling augmentation, what are the results of using a single encoder with average loss and worst-view loss, respectively? Compared to a dual encoder with a matched parameter count, does complementary pretraining still have a consistent advantage?

We thank the reviewer for this important point. As suggested, each comparison below changes one factor and keeps everything else identical (sampling variations with K=4 wherever used, downstream training budget, model selection).

Sampling variations alone do not explain the gain. Table 3 adds the components in one order, so its steps do not isolate them; even there, complementary pretraining is the largest step on eICU (+2.15 AUROC). Fixing PRISM's dual encoder, we varied the two components independently:

Complementary pretrainingSampling variations (worst-view)HAR-C (acc)eICU (AUROC)
––69.00±1.1579.93±0.13
✓–72.02±0.4381.87±0.07
–✓79.98±0.2881.46±0.08
✓✓ (PRISM)80.57±0.1982.47±0.04

Complementary pretraining helps without (+3.03±1.07 / +1.94±0.12 on HAR-C / eICU) and with sampling variations (+0.58±0.24 / +1.00±0.09; all p-value≤0.04), and neither component alone reaches PRISM; the strongest baseline on HAR-C, MLDG, trained with the same variations stays 2.12±0.21 below PRISM there (p-value=4e-6). As noted in Section 4.3, the contributions follow the type of shift: sampling variations contribute more on HAR-C, whose only shift between the training and test domains is the sampling pattern shift by construction, and complementary pretraining on eICU, where sampling pattern shifts are entangled with other real-world variation. Where the sampling cue misleads, complementary pretraining matters most: on the shortcut benchmark with the same sampling variations, it raises bias-conflicting accuracy by 10.81±1.26 points (p-value=1e-5; W2).

Single encoder with average and worst-view loss. Trained with the same sampling variations and no pretraining, it shows what sampling augmentation alone achieves; with either loss it stays below PRISM (+2.10±0.23 on HAR-C and +1.04±0.10 on eICU over the worst-view variant; p-value≤6e-6):

ModelHAR-C (acc)eICU (AUROC)
Single encoder (1×), average loss77.47±0.7180.49±0.10
Single encoder (1×), worst-view loss78.46±0.2681.43±0.07
PRISM (dual encoder, 2.01×)80.57±0.1982.47±0.04

Complementary pretraining at matched or larger capacity. All variants use the worst-view loss; sizes are relative to one CNN encoder (HAR-C / eICU for the larger single encoder):

Encoder (size)PretrainingHAR-C (acc)eICU (AUROC)
Single, larger (2.48× / 4.15×)–78.77±0.2281.43±0.12
Single, larger (2.48× / 4.15×)value + sampling pattern reconstruction79.34±0.4881.98±0.05
Dual (2.01×)–79.98±0.2881.46±0.08
Dual (2.01×), PRISMcomplementary80.57±0.1982.47±0.04

PRISM outperforms the same-size dual encoder without pretraining (+0.58±0.24 / +1.00±0.09), a larger single encoder (+1.80±0.15 / +1.04±0.13), and one trained with both reconstruction objectives (+1.23±0.37 / +0.49±0.07; all p-value≤0.04): the advantage comes not from capacity or the objectives alone, but from learning them in separate encoders. We have added these results as Table A11 and to Section 4.3.

[W2 & Q4]

There is insufficient evidence for PRISM’s mechanism to reduce reliance on sampling shortcuts. In Figure 4, PRISM’s mask-following rate is similar to that of ERM, making it unclear whether the improvement comes from stronger utilization of numerical features or from reduced dependence on sampling shortcuts.

Analysis of prediction changes when only changing the sampling mode while keeping the base sequence fixed, and comparing it with a single encoder under the same augmentation to more directly test sampling dependence?

We thank the reviewer for this question. In the Figure 4 shortcut benchmark, each class is observed with its own sampling pattern with probability ρ=0.9 in training, a strongly predictive cue that the other methods follow (mask-following rate 0.416–0.526). PRISM's rate is the lowest (0.405) but close to ERM's (0.435), as PRISM is not designed to discard sampling pattern information, i.e., to be invariant to sampling patterns (2NXh, W4), which can be regarded as informative in this biased dataset.

As suggested, we fix the base sequence and change only the sampling: for each HAR-C test window and each of the 8 conditions, we draw two independent masks of the same type and measure redraw flip, how often the prediction changes between them.

Redraw flip, all 8 (%)Redraw flip, Fixed-/Rand-Feat. (%)
ERM33.7±1.655.4±3.2
MLDG (best on HAR-C)28.4±1.055.8±3.0
ManyDG (best on shortcut benchmark)30.4±1.053.6±2.9
Single encoder, average loss (A)19.9±0.436.2±1.3
Single encoder, worst-view loss (B)17.6±0.232.8±0.9
Dual encoder, no pretraining (E)17.1±0.231.7±0.9
PRISM16.9±0.330.6±0.8

Keeping sampling information, PRISM naturally still responds to the mask, but flips far less often than ERM and the best baselines (16.9% vs. 28.4–33.7%, all p-value≤1e-6) and, under the feature-wise sampling pattern shifts (Fixed-Feat. and Rand-Feat.), is significantly more stable than a single encoder with the same augmentation (B; 32.8±0.9% → 30.6±0.8%, p-value=5e-4). We have added these results to Section 4.2 (Table A13).

The use of feature values is tested where the sampling cue misleads (shortcut benchmark, ρ=0.9; Tables A6 and A16), now also with B and E trained with the same sampling variations:

Bias-alignedUnbiasedBias-conflictingMask-following rate
ERM93.99±2.4458.00±0.6942.59±1.860.435±0.019
ManyDG (best baseline)93.55±0.7959.17±0.8144.87±1.260.416±0.017
Single encoder, worst-view loss (B)99.00±0.1253.26±1.2832.86±1.980.556±0.016
Dual encoder, no pretraining (E)99.00±0.1257.80±0.9838.30±1.300.498±0.011
PRISM92.29±2.5964.98±0.8349.11±0.740.405±0.015

PRISM is the most accurate of all methods evaluated when the sampling cue is uninformative or misleading, showing that it uses the feature information. With the architecture, sampling variations and worst-view objective fixed (E vs. PRISM), complementary pretraining improves bias-conflicting accuracy by 10.81±1.26 points (p-value=1e-5) and reduces predictions that follow the conflicting sampling cue by 9.3±1.5 points (mask-following rate 0.498 → 0.405, p-value=1e-4); without it, models trained with the same sampling variations follow the sampling cue more (B, E). We have added these results to Section 4.2 (Table A16).

In conclusion, the improvement comes from both effects the reviewer names, as intended by design: PRISM keeps sampling pattern information but is less sensitive to it, and when the sampling cue misleads, complementary pretraining makes it follow the cue less and predict the true class from the feature information instead (9.3 of its 10.81-point gain in bias-conflicting accuracy comes from fewer cue-following predictions).

[W3]

HAR-C controls the number of observations but does not rule out the effects of task-relevant feature changes. Different time periods and sensors may contain varying levels of class-discriminative information; therefore, the same number of observations does not necessarily imply the same prediction difficulty. The performance drop in HAR-C may result from both the failure of sampling shortcuts and the absence of key discriminative information.

We thank the reviewer for this point. Indeed, the same number of observations does not imply the same difficulty, and this diversity of HAR-C's shifts is intended. Existing evaluations remove a growing fraction of observations at test time [1, 2], so lost information and sampling pattern change grow together; HAR-C instead fixes the number of observations and moves them to different time periods and sensors under 8 sampling conditions, which naturally retain different amounts of class-discriminative information, just as ImageNet-C corruption types [3] differ in difficulty.

We propose a matched-pattern control to distinguish this difficulty from the sampling pattern shift between the training and test domains: we train ERM on each of the 8 sampling patterns and test it on all 8 (Table A12; ERM is retrained for this control, with Fixed-Feat. dropping the same channels in training and test). Training on the test pattern itself gives the accuracy reachable without a shift, so the remaining gap of the model trained on Random is due to the shift:

AlgorithmRandomRegularDesyncFixed-Feat.Rand-Feat.FirstLastMidAvg.
ERM, trained on Random80.6±0.582.1±0.969.3±1.159.2±3.554.2±0.567.6±1.855.6±3.163.9±2.366.6±0.7
ERM, trained on test pattern80.6±0.586.4±0.380.1±0.476.9±1.064.7±0.481.2±0.481.1±0.480.7±0.379.0±0.1
PRISM, trained on Random84.1±0.286.5±0.282.8±0.374.9±1.374.7±0.281.3±0.279.2±0.481.0±0.280.6±0.2

Of the 14.0±0.7-point average drop (80.6 → 66.6), 12.4±0.6 points are recovered by training on the test pattern (66.6 → 79.0), so most of it comes from the shift, which accounts for the larger part in every other condition with a drop except Rand-Feat. (15.8±0.7 of 26.4±0.9 points due to its difficulty). We now state this in Section 2.2 and Appendix D. Notably, PRISM, trained only on Random, is on average even more accurate than ERM trained on the test pattern (80.6±0.2 vs. 79.0±0.1, p-value=4e-5), by up to 10.0±0.3 points under Rand-Feat., the hardest condition.

[1] Zhang et al. "Graph-Guided Network for Irregularly Sampled Multivariate Time Series." ICLR 2022.
[2] Li et al. "Time Series as Images: Vision Transformer for Irregularly Sampled Time Series." NeurIPS 2023.
[3] Hendrycks and Dietterich. "Benchmarking Neural Network Robustness to Common Corruptions and Perturbations." ICLR 2019.


[Q2]

Table A8 marks K=1 as 'no sampling change.' According to the formula, K=1 may still apply one random transformation to the input. Does this ablation actually use the original input or a randomly augmented view?

Thank you for catching this ambiguity. The K=1 row uses the original input: no sampling variation is generated and the pretrained model is fine-tuned with plain ERM on x, making it the true no-augmentation ablation (72.02±0.43 on HAR-C, 81.87±0.07 on eICU; Table 3), now relabeled "K=1 (original input, no sampling variations)" in Tables A8 and A9.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We thank all reviewers for their constructive feedback and engagement in the discussion.


All main concerns are addressed: 6Qi7 now finds that the paper can be accepted, 6Qi7 and 7rtk confirm that our responses address their main concerns, and all three reviewers acknowledge that the revisions improved the paper.

  • 6Qi7: the modifications in the paper "address my major concerns", and "the paper can be accepted".
  • 7rtk: "The factorial ablations, the matched-pattern control and the redraw-flip analysis address my main concerns", and "the rebuttal has strengthened the paper". The remaining point, robustness to shifts outside the augmentation family, is addressed in our follow-up (last item below).
  • 2NXh: improving the clarity and adding more analysis "improved this work".

Reviewers also found the problem well motivated and the research question valuable (7rtk), the paper well written (6Qi7), the analysis valuable to the community (2NXh), and PRISM's results strong across datasets, backbones and ablations (7rtk, 2NXh).


How we addressed the main concerns (new experiments: 10 seeds; mean±s.e.; paired t-tests)

  • Does complementary pretraining add beyond sampling augmentation? (7rtk, 2NXh) Yes (Table A11). With the same dual encoder and sampling variations, it adds +0.58±0.24 (HAR-C) and +1.00±0.09 AUROC (eICU), and PRISM outperforms a larger single encoder trained with both reconstruction objectives (+1.23±0.37 / +0.49±0.07). MLDG given the same sampling variations still trails PRISM by 2.12±0.21 on HAR-C.
  • Does PRISM reduce reliance on sampling shortcuts? (7rtk, 2NXh) Yes. Under the same sampling variations, complementary pretraining raises bias-conflicting accuracy from 38.30±1.30 to 49.11±0.74 and lowers the mask-following rate from 0.498 to 0.405 (p-value≤1e-4); in the redraw-flip analysis, redrawing only the mask flips PRISM's predictions half as often as ERM's (16.9% vs. 33.7%).
  • Does the HAR-C drop come from the sampling pattern shift or from less discriminative observations? (7rtk, 6Qi7) Mostly from the shift: in the matched-pattern control, training on the test pattern recovers 12.4±0.6 of the 14.0±0.7-point average drop. PRISM, trained only on Random, is on average even more accurate than ERM trained on the test pattern (80.6 vs. 79.0).
  • Do the Section 2 findings hold beyond HAR? (6Qi7) Yes, the controlled analyses replicate on PAM, and PRISM leads on PAM-C.
  • Recent methods and MLDG (6Qi7). MEDRO (NeurIPS'25) and Hierarchical DRO (ICLR'26) stay below MLDG and far below PRISM. MLDG is close to PRISM only where the shift is absent or mild; on average, PRISM leads by 7.87±0.51 on HAR-C.
  • Clarity and details (2NXh, 7rtk). Sections 2 and 3 are revised (below); protocol details are in our response to 7rtk.
  • Shifts outside the augmentation family (7rtk, follow-up). The sampling variations drop random time points and features, so they resemble only the feature-wise shifts. Under the other five HAR-C shifts (Regular, Desync, First, Last, Mid), PRISM outperforms MLDG by 5.73±0.39 points and, trained only on Random, is as accurate as ERM trained on each test pattern (82.16±0.15 vs. 81.89±0.18); MLDG given the same variations still trails PRISM by 1.72±0.16 (p-value=2e-6). PRISM also leads on these shifts in PAM-C and in all four target regions of eICU, where the shifts are real.

Revisions (in blue in the revised paper)

  • Section 2 now opens with the problem setup and our scope (classification), and Section 2.2 defines the full input, the mask-only input, bias-conflicting accuracy and the mask-following rate; its logical flow and the motivation of Section 3.1 are clarified.
  • New Tables A11–A16 for the experiments above, MEDRO and Hierarchical DRO in Tables 1, 2 and A7, and probe details in Appendix D.
  • Related work extended with the HAR study suggested by 6Qi7 (Section 2.1) and recent works (Appendix B).
  • All code is in the restored repository: https://anonymous.4open.science/r/PRISM-8C66.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.