acceptodds
Under review as a conference paper at ICLR 2027

Matching Scales to Patterns: Soft-Segmented Multi-scale Experts for Irregular Multivariate Time Series Forecasting

Abstract

Irregular multivariate time series (IMTS) are pervasive in real-world scenarios, characterized by uneven sampling intervals and misaligned observations across variables. Their temporal patterns can vary substantially across samples and periods in aspects such as fluctuation dynamics and changes in sampling density. Distinct aspects of these patterns can be captured at multiple scales, while their variation across samples and evolution over time may call for different scale combinations. However, existing multi-scale IMTS forecasting methods typically apply a fixed scale configuration to all samples and time periods, limiting their ability to adapt scales to varying temporal patterns. To address this limitation, we propose SoftScale, an adaptive soft-segmented multi-scale experts network for IMTS forecasting. SoftScale first employs data-dependent time cursors to construct sample-adaptive soft segments from raw IMTS. For each soft segment, a multi-scale experts network then captures temporal and inter-variable dependencies at each scale. To integrate these dependencies, an MoE router adaptively selects the most suitable scale combination for each soft segment and combines the corresponding scale-specific representations. In this way, SoftScale enables adaptive multi-scale modeling across different samples and periods, leading to more accurate IMTS forecasting. Experiments on four real-world datasets demonstrate that SoftScale outperforms existing state-of-the-art baselines. Our code is available at [https://anonymous.4open.science/r/SoftScale/](https://anonymous.4open.science/r/SoftScale/).

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series, Multi-scale, Multivariate Time Series Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ELJ1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SoftScale for irregular multivariate time-series forecasting. Adaptive time cursors construct overlapping soft segments from observation timestamps. Scale-specific temporal and inter-variable attention produce representations that a segment-conditioned router combines before predicting queried variable–time pairs. The intended contribution is to adapt scale combinations across samples and temporal regions instead of using one fixed configuration. Evaluation on four benchmarks includes comparisons with established IMTS methods, five-seed results, component ablations, routing perturbations, and efficiency analysis.

Critical strengths
  1. A relevant adaptation of multi-scale modeling to irregular observations. The combination of soft temporal assignment and segment-specific scale fusion addresses a concrete limitation of fixed temporal aggregation in IMTS. It connects sampling structure, temporal aggregation, and inter-variable modeling in one forecasting architecture. This is useful for settings in which observation density changes substantially across patients or periods.

  2. Several complementary experiments support practical usefulness. Table 1 reports improvements over strong IMTS baselines across the four default settings, with five-seed uncertainty estimates. Tables 2–3 go beyond the main comparison by removing components and perturbing routing at inference. For example, PhysioNet MSE increases from 4.57 to 4.72 (both ×10⁻³) when routing weights are exchanged across samples. These experiments provide evidence that the learned routing is useful, although they do not establish every proposed interpretation of it.

Critical weaknesses
  1. The interpretation of adaptation to value dynamics is stronger than the routing equations support. In Sections 4.1 and 4.2.3, cursor drift and the routing input are functions of timestamps and learned embeddings, without conditioning on observed values. Even with the noise term in Eq. (10), both cW_r and Softplus(cW_noise) are computed from c, so two inputs with the same observation times and variable identities receive identically distributed routing weights even if their fluctuation patterns differ. The experts themselves do consume values, so this does not invalidate forecasting; it limits the claim that the router selects scales in response to value dynamics. This distinction matters because matching scales to changing temporal patterns is the central motivation.

  2. The incremental benefit of the central mechanisms varies considerably across datasets. Table 2 includes relevant simpler alternatives, such as unweighted summation (w/o MoE) and same-scale experts (w/o Multi-Scale). The gains from ACSS are clear on PhysioNet (MSE 4.57 ± 0.07 vs. 4.78 ± 0.07, ×10⁻³) and MIMIC-III (MAE 6.85 ± 0.08 vs. 7.12 ± 0.05, ×10⁻²), but small on Human Activity (3.19 ± 0.02 vs. 3.22 ± 0.04, ×10⁻³) and USHCN (4.98 ± 0.07 vs. 5.03 ± 0.05, ×10⁻¹). The USHCN main-table margin over the closest IMTS baseline is likewise small (4.98 ± 0.07 vs. GraFITi 5.01 ± 0.03). Because only per-variant means and standard deviations are reported, the reliability of the smaller differences cannot be judged from paired comparisons. The issue is the strength of evidence for the specific architectural increment in some settings, rather than whether every individual result is state of the art.

Questions for authors
  1. Clarification of the main mechanism: Is the timestamp-only dependence described above the intended implementation? If so, which claims concern adaptation to sampling patterns, and which concern adaptation to value dynamics? If observed values enter the router elsewhere, please identify that path.

  2. Targeted validation of the architectural increment: Can you provide the paired seed-level differences, or corresponding uncertainty estimates, for the ACSS and routing ablations? Please also report the parameter counts of the ablation variants and the number of epochs actually trained under the shared early-stopping rule; this would help determine how much of the improvement is specifically due to segmentation and adaptive fusion.

6 comments

Responses to Reviewer ELJ1 (1/2)

W1/Q1: Clarification of the main mechanism

We thank the reviewer for raising this point. We agree that the router receives only timestamp-derived representations at inference and therefore cannot directly identify value dynamics on its own. However, value dynamics can still help shape the routing function and influence scale weights. Specifically, the router allocates scale weights based on irregular sampling patterns, while the scale experts model observed values at different scales. During end-to-end training, value dynamics implicitly shape the learned routing function through the experts’ outputs and the forecasting loss. This allows the router to learn which scale mixtures work well for a given sampling context, taking into account how well the experts capture value dynamics during training. The following evidence supports this interpretation.

(1) Relevance of timestamp-conditioned routing. Table 3 provides quantitative evidence that the learned scale mixtures are meaningful for forecasting. Figure 3 complements this result with a qualitative visualization of how the learned scale mixtures differ across samples and soft segments.

(2) Effect of value dynamics on router training. On PhysioNet, we perturbed only the historical values while keeping the observation timestamps and model parameters fixed. We reduced or increased the historical standard deviation by 25% while preserving the mean, or added an upward or downward linear trend using centered, range-normalized timestamps and a coefficient of 0.2 times the original standard deviation. Across 18,441 eligible histories, we compared the forecasting-loss gradients with respect to the routing logits before and after each intervention.

Value interventionGradient change ≥ 10%Gradient change ≥ 25%Gradient–Oracle agreement
Lower volatility62.3%33.7%88.1%
Higher volatility62.9%34.9%87.3%
Downward trend52.6%22.5%88.1%
Upward trend51.8%21.7%88.3%

Gradient–Oracle agreement measures how often the expert with the lowest routing-logit gradient also yields the lowest local MSE when given extra weight. We check this preference by giving each expert more weight in turn while keeping the expert outputs fixed. For each trial, we keep 80% of the original mixture and give the remaining 20% to the expert being tested. The expert whose trial yields the lowest MSE is the oracle choice. Agreement was 87.3%–88.3% across the four interventions. The gradient changes show that value dynamics affect the signal used to train the router, and the high agreement shows that this signal tends to favor the scales that best serve the perturbed sequence. Through joint optimization, SoftScale learns to match scales to temporal patterns, with the router conditioned on sampling patterns and implicitly shaped by value dynamics.

(3) Directly adding observed values to the router. To test whether feeding observed values directly into the router improves forecasting, we evaluated two variants.

  • Full-history Values adds encoded observed-value features together with their mean and standard deviation, computed over each variable’s full observed history.
  • Segment-wise Values weights encoded observed-value features by the ACSS assignments and additionally includes the corresponding assignment-weighted mean and standard deviation for each soft segment.

Both variants have identical parameter counts, averaging 1.64% more parameters than SoftScale. We report MSE as mean ± sample standard deviation over five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
+ Full-history Values4.99 ± 0.064.68 ± 0.061.77 ± 0.113.22 ± 0.04
+ Segment-wise Values5.00 ± 0.054.74 ± 0.171.79 ± 0.083.22 ± 0.04

Analysis. Neither value-aware variant improved mean MSE. The gaps were larger on the clinical datasets, PhysioNet and MIMIC-III. Feeding value features directly into the router may expose it to more complex, variable, and sometimes less reliable signals, making scale allocation more sensitive to transient fluctuations. On Human Activity and USHCN, the experts may already capture recurring motion and seasonal patterns, limiting the benefit of adding value features to the router.

We will clarify that SoftScale adapts its scale weights to the sampling context at inference, with value dynamics shaping the learned routing function during training. In this way, SoftScale matches scales to temporal patterns through scale adaptation that is explicitly conditioned on sampling patterns and implicitly shaped by value dynamics.

Responses to Reviewer ELJ1 (2/2)

W2/Q2: Targeted validation of the architectural increment

Why the gains vary across datasets. In clinical time series, sampling density and value dynamics vary across patients and stages of care. Scale adaptation can therefore offer greater benefits when local patterns and observation density differ across segments, consistent with the larger gains on PhysioNet and MIMIC-III. Human Activity exhibits motion continuity and repeated actions, while USHCN exhibits seasonal patterns. These more regular and structured patterns may lead to more consistent scale preferences across segments, helping explain why the gains are smaller on these two datasets. SoftScale nevertheless achieves lower mean MSE on both datasets in the controlled comparisons below.

(1) Dataset-level routing interventions. We applied paired significance tests across five training seeds to assess the MSE increases from the routing interventions reported in Table 3 of the manuscript on the PhysioNet test set.

Routing intervention95% CI for MSE change (×10⁻³)Seeds with higher MSE
Sample Shuffle[0.0443, 0.2409]5/5
Both Segments Expert Shift[0.0578, 0.3381]5/5
Segment 0 Expert Shift[−0.0130, 0.1050]4/5
Segment 1 Expert Shift[0.0075, 0.1223]5/5

Three of the four interventions significantly increased MSE in separate two-sided paired tt-tests across five training seeds (p<0.05p<0.05).

(2) Paired ablations. We compare three settings using the same five random seeds.

  • SoftScale: uses both ACSS and adaptive routing.
  • w/o ACSS: removes adaptive soft segmentation while retaining the multiscale experts and adaptive routing.
  • w/o MoE: retains ACSS and the multiscale experts but bypasses adaptive routing and sums the expert outputs without routing weights.

The table reports the mean and sample standard deviation of paired MSE and MAE differences (SoftScale minus the ablation). Wins counts the seeds on which SoftScale obtains a lower error for the corresponding metric. Parameter counts and training epochs are shown as SoftScale / ablation. All settings use the same early-stopping rule with a patience of 20 epochs.

DatasetAblationΔ\DeltaMSE (mean ± SD)MSE winsΔ\DeltaMAE (mean ± SD)MAE winsParametersTraining epochs (mean ± SD)
MIMIC-IIIw/o ACSS(−8.20±6.30)×10−4(-8.20\pm6.30)\times10^{-4}5/5(−2.696±0.643)×10−3(-2.696\pm0.643)\times10^{-3}5/5708,389 / 707,55541.2 ± 5.5 / 66.6 ± 24.3
MIMIC-IIIw/o MoE(−7.70±5.33)×10−4(-7.70\pm5.33)\times10^{-4}5/5(−1.230±2.576)×10−3(-1.230\pm2.576)\times10^{-3}4/5708,389 / 708,38941.2 ± 5.5 / 46.0 ± 9.9
PhysioNetw/o ACSS(−2.06±0.79)×10−4(-2.06\pm0.79)\times10^{-4}5/5(−8.42±6.53)×10−4(-8.42\pm6.53)\times10^{-4}5/51,070,208 / 1,069,48447.6 ± 7.3 / 57.8 ± 22.6
PhysioNetw/o MoE(−1.54±1.38)×10−4(-1.54\pm1.38)\times10^{-4}5/5(−1.96±5.44)×10−4(-1.96\pm5.44)\times10^{-4}3/51,070,208 / 1,070,20847.6 ± 7.3 / 53.8 ± 10.4
Human Activityw/o ACSS(−3.8±3.1)×10−5(-3.8\pm3.1)\times10^{-5}5/5(−3.76±4.31)×10−4(-3.76\pm4.31)\times10^{-4}3/5222,857 / 222,51188.8 ± 6.9 / 112.6 ± 31.3
Human Activityw/o MoE(−4.6±2.7)×10−5(-4.6\pm2.7)\times10^{-5}5/5(−4.26±3.85)×10−4(-4.26\pm3.85)\times10^{-4}4/5222,857 / 222,85788.8 ± 6.9 / 78.2 ± 7.4
USHCNw/o ACSS(−5.324±10.835)×10−3(-5.324\pm10.835)\times10^{-3}3/5(−1.2564±1.1365)×10−2(-1.2564\pm1.1365)\times10^{-2}5/5157,748 / 157,41681.2 ± 4.9 / 101.8 ± 16.3
USHCNw/o MoE(−6.176±7.505)×10−3(-6.176\pm7.505)\times10^{-3}4/5(−4.140±12.392)×10−3(-4.140\pm12.392)\times10^{-3}3/5157,748 / 157,74881.2 ± 4.9 / 88.8 ± 11.2

SoftScale wins 90% (36/40) of paired comparisons against w/o ACSS and 82.5% (33/40) against w/o MoE. The w/o MoE variant has the same instantiated parameter count as SoftScale, and removing ACSS reduces the count by only 332–834 parameters. SoftScale trains for fewer epochs on average in seven of the eight comparisons. These comparisons suggest that differences in parameter count or training duration are unlikely to explain SoftScale’s gains.

(3) Learned static versus adaptive fusion. To test whether learning a single fixed set of weights can account for SoftScale’s gains, we additionally evaluated a Global Static baseline on all four datasets. Global Static retains the scale experts and learns one weight vector shared across all samples, variables, and soft segments. The table reports MSE as mean ± sample standard deviation across five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
Global Static5.10 ± 0.044.76 ± 0.191.78 ± 0.103.27 ± 0.03

SoftScale achieves lower mean MSE on all four datasets. Together with the routing interventions, this comparison supports the forecasting value of adapting scale mixtures across samples and segments.

We hope these responses address your concerns. Thank you for your constructive feedback.

Thanks for the detailed response and new experiments.

The paired ablations and the Global Static baseline make a much stronger case for ACSS and adaptive scale fusion. Even though some of the gains are still marginal, this resolves my second concern. However, on my first concern, your response does not change the core fact: while the gradient analysis shows that value dynamics influence router training, inference-time routing still depends solely on sampling patterns rather than observed values—and the value-aware ablations do not change this. This directly contradicts the paper's claim that SoftScale adapts routing decisions to value dynamics, which undermines the stated contribution of the paper.

Thanks again for the thorough rebuttal.

Reviewer_pSJsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper proposes SoftScale for irregular multivariate time series forecasting. It uses time cursors to construct soft segments, applies multi-scale experts to extract scale-specific representations, and uses an MoE router to combine the scale outputs. The paper reports competitive results on four datasets and includes ablation, routing perturbation, efficiency, and scale-configuration analyses.

Critical strengths
  • The motivation is reasonable. It is meaningful to ask how adaptive scales can handle irregular sampling.
  • The paper compares against a range of baselines and shows empirical advantages.
  • The paper does not only report main results. It also includes ablation, routing perturbation, and efficiency analyses, which help explain why the method might work.
Critical weaknesses
  • The method still uses a predefined scale pool and combines all scale outputs through weighted aggregation. No scale is explicitly discarded, and the final output is still a fusion of multiple scales. From this perspective, the method does not appear fundamentally different from other adaptive multi-scale approaches.
  • MoE is used as a simple weighted fusion mechanism, without a novel scale-selection mechanism. This also suggests that scale selection still relies on passive learning. In addition, it is unclear what replaces scale fusion in the w/o MoE variant.

The code repository is not accessible.

Questions for authors
  1. If the scale pool remains predefined and no scale is explicitly discarded, what is the essential difference between SoftScale and other adaptive multi-scale methods?
  2. Can the authors provide a direct comparison under the same scale pool and parameter budget between fixed multi-scale fusion and soft multi-scale fusion?
  3. What exactly replaces scale fusion in the w/o MoE variant, and why is the MoE fusion considered a methodological contribution rather than a simple weighted combination?
  4. How do different scale sizes, rather than only the number of scales, affect training time, inference time, parameter count, and memory?
  5. Are future query timestamps and variable indicators known at deployment time, and do all baselines use the same information? Will the code be made publicly available?
10 comments

Responses to Reviewer pSJs (1/2)

W1/Q1: Differences from existing adaptive multiscale methods and the methodological contribution of MoE fusion

(1) Difference from existing adaptive multiscale methods. In IMTS, value dynamics and sampling density vary across samples and periods. Pathformer [1] selects scale pathways based on the input sequence. By contrast, SoftScale adapts scales at the level of sample-adaptive soft segments in two ways.

  1. Scale adaptation across samples. SoftScale uses the original observation timestamps of each sample to construct soft segments with adaptive time cursors. It then learns scale mixtures conditioned on these segments, allowing it to adapt to the irregular temporal patterns of different samples.
  2. Scale adaptation within a sample through soft segmentation. SoftScale learns a separate scale mixture for each soft segment within a sample. When temporal patterns change gradually, observations near a segment boundary can contribute to multiple segments with different weights. This reduces the separation of observations with similar patterns that hard boundaries can impose. Even with predefined expert scales and all experts contributing to the fusion, each expert's contribution can be adjusted to the temporal pattern of each segment.

(2) How MoE implements segment-level scale adaptation. Each expert corresponds to a specific candidate scale and extracts features from the same soft segment at that scale. The temporal representation of the segment guides both feature extraction at each scale and the router’s allocation of weights across scales. The methodological contribution of MoE lies in jointly learning scale-specific temporal dependencies within soft segments and adapting their relative contributions across samples and segments. For example, one segment may rely more heavily on fine-scale information, while another may give greater weight to coarser patterns. A weighted sum alone provides no mechanism for learning how these weights should vary with the temporal context of each segment.

(3) Why use dense routing? A single soft segment may contain both short-term fluctuations and slower trends. Dense routing preserves all scale-specific views and adjusts their contributions through continuous weights, allowing complementary information from multiple scales to contribute to the prediction. In sparse, asynchronous IMTS, limited observations may introduce uncertainty into local scale selection. To examine whether explicitly discarding a scale benefits forecasting, we compared SoftScale with Top-2 fusion on all four datasets. The table reports the mean and sample standard deviation of MSE across five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
Top-2 fusion5.06 ± 0.074.92 ± 0.101.74 ± 0.023.23 ± 0.05

Dense routing achieves lower mean MSE on all four datasets, with larger relative gaps on PhysioNet and MIMIC-III. One possible explanation is that clinical observations are often sparse and asynchronous, and abrupt local changes can coexist with slower variations. Discarding a scale may therefore remove complementary temporal information. Motion continuity and repeated actions in Human Activity, together with the seasonality of weather variables in USHCN, may yield more regular dominant patterns and reduce the impact of discarding a scale.

W2/Q3: What exactly replaces scale fusion in the w/o MoE variant?

Implementation of the variant. The w/o MoE variant retains the same scale pool, number of experts, and expert architectures while removing the input-conditioned router. The complete model uses

Hi,v,p=∑s=1Kgi,v,p,sEs(i,v,p),\mathbf H_{i,v,p}=\sum_{s=1}^{K}g_{i,v,p,s}\mathbf E_s(i,v,p),

whereas w/o MoE directly sums the same expert representations without routing weights:

Hi,v,pw/o MoE=∑s=1KEs(i,v,p).\mathbf H^{\mathrm{w/o\ MoE}}_{i,v,p}=\sum_{s=1}^{K}\mathbf E_s(i,v,p).

The variant therefore retains multiscale representation learning and applies the same unweighted aggregation to every sample, variable, and soft segment. In the revised manuscript, we will consider renaming this variant from w/o MoE to w/o router to clarify that the scale experts are retained.

W3/Q5: Code access

Code availability. We apologize that the code was not publicly accessible during the early review period due to an error in the scheduled release settings. We corrected this on October 22. The code is now publicly available and will remain accessible throughout the rebuttal period.

References

[1] Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting, Peng Chen et al., ICLR 2024.

Responses to Reviewer pSJs (2/2)

Q2: Fixed versus adaptive multiscale fusion under the same scale pool and parameter budget

The w/o MoE ablation in Table 2 already provides a controlled comparison with the same scale pool and expert parameter budget. It retains one expert per candidate scale and sums the three scale-specific outputs with equal weights. To test whether learning a single fixed set of weights can account for SoftScale’s gains, we additionally evaluated a Global Static baseline on all four datasets.

Experimental setup. In Global Static, all samples, variables, and soft segments share the same learnable vector of expert weights. The table reports MSE as mean ± sample standard deviation across five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
Global Static5.10 ± 0.044.76 ± 0.191.78 ± 0.103.27 ± 0.03

Results and analysis. SoftScale achieves lower average MSE on all four datasets, with larger gains on PhysioNet and MIMIC-III. One possible explanation is the substantial heterogeneity across patients in clinical time series. Sampling density and value dynamics can also vary over a patient’s clinical course, creating different scale requirements across samples and time segments. The seasonality of weather variables in USHCN, together with motion continuity and repetitive action structures in Human Activity, may lead to more similar scale preferences across samples, narrowing the performance gap between globally shared weights and dynamic scale allocation.

Together with the routing intervention results reported in Table 3 of the manuscript, this experiment further supports the contribution of sample- and segment-specific scale allocation to forecasting performance.

Q4: Effect of expert scale size on compute cost

(1) Experimental setup. We fixed the number of scale experts to three on PhysioNet and maintained a 4:2:14:2:1 ratio among their sizes. We tested six configurations: [32,16,8], [64,32,16], [96,48,24], [128,64,32], [192,96,48], and [256,128,64]. We kept the benchmark setup fixed across configurations on an RTX 3090. We fitted power-law exponents α\alpha against the sum of the three scale sizes to characterize how each cost grew over the tested range.

Varied factorTested rangeTraining time (ms/step)Inference time (ms/batch)ParametersPeak memory: Train / Inference (MB)α\alpha: Train / Infer / Params / Train mem / Infer mem
Scale size (3 experts)[32,16,8] → [256,128,64], S:56→448S:56\to448 (8×)33.04±2.66→46.16±3.9833.04\pm2.66\to46.16\pm3.98 (1.40×)13.67±0.59→19.61±0.7013.67\pm0.59\to19.61\pm0.70 (1.43×)0.586M → 1.715M (2.93×)1898.4 → 2526.2 (1.33×) / 794.1 → 906.3 (1.14×)0.16 / 0.17 / 0.52 / 0.14 / 0.05

(2) Interpretation. Let SS be the sum of the three expert scale sizes. The model has 425,088 parameters independent of SS and adds 2,880 parameters for each unit increase in SS. Over the tested range (S=56S=56 to 448448), the fitted training time is 31.27+0.03335S31.27+0.03335S ms per step, inference time is 12.93+0.01515S12.93+0.01515S ms per batch, and peak training memory is 1807.54+1.601S1807.54+1.601S MB. Only the terms proportional to SS grow as the scales increase. Peak inference memory rose from 794.1 to 906.3 MB, an increase of 112.2 MB (14.1%), as SS increased eightfold.

We will include the theoretical complexity analysis with respect to expert size and the results of this experiment in the appendix of the revised manuscript.

Q5: Future query inputs and baseline fairness

Future query timestamps and variable indicators. At deployment, the model receives the requested future timestamps as input, and variable identities follow a predefined schema. Given the observed history, SoftScale predicts all variables at each requested future timestamp. The future observation mask marks which targets have ground-truth values and is used to exclude missing targets from the training loss and evaluation metrics. The forecasting network receives no future observation mask.

Baseline information and fairness. All methods are evaluated on the same targets. TimesNet, TimeMixer, TimeFilter, ModernTCN, PatchTST, and MSGNet are regular time-series baselines that predict a fixed number of steps without using explicit future query timestamps. All other regular baselines and all irregular baselines use future query timestamps or their time features. All methods use variable identities through fixed channels, embeddings, or variable nodes. We follow the official settings for each baseline, including its use of future query timestamps and variable indicators.

We hope these responses address the concerns and respectfully request reconsideration.

I thank the authors for the engineering-oriented strengthening regarding Q2–Q5, especially the clarification of the w/o MoE variant, the Global Static comparison, the scale-size cost analysis, and the code availability fix; I acknowledge that these efforts improve the paper's transparency and experimental completeness.

However, my core concern remains Q1: the nature of the method and the positioning of its contribution. The combination of a predefined scale pool and dense weighted fusion is more of an engineering combination and does not introduce a new scale selection mechanism; the additional experiments demonstrate the utility of dynamic weighting, rather than the novelty of the method or its fundamental distinction from other adaptive multi-scale methods. In addition, the authors acknowledge in the rebuttal that the router relies solely on timestamps at inference. This helps clarify the paper's related statements, but it further weakens the original motivation of “matching scales to patterns”—because “patterns” in that motivation should include value dynamics. Therefore, for the current version and rebuttal, my core concern has not been fully addressed by the experiments or analysis.

Based on the above, I am willing to raise the Soundness score accordingly, but I will maintain the overall rating for now. I thank the authors again for their rebuttal.

Reviewer_WMx8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SoftScale, an adaptive multi-scale modeling framework for irregular multivariate time series forecasting. Its key contributions include:

  1. An Adaptive Cursor-based Soft Segment module that constructs sample-adaptive soft segments using data-dependent time cursors.
  2. A multi-scale experts network that captures temporal and inter-variable dependencies at different scales.
  3. An MoE router that adaptively combines scale-specific representations. Experiments on the four IMTS benchmarks demonstrate consistent improvements over 22 baselines and achieve state-of-the-art performance.
Critical strengths
  1. The method description and presentation are detailed and clear, with explicit mathematical formulations and well-organized figures. The authors also provide code to aid reproducibility.

  2. The experimental scope is commendable and comprehensive, with systematic ablation studies.

Critical weaknesses
  1. The implementation of the routing mechanism seems inconsistent with the stated motivation. Qadapt=Qinit+λ⋅fdrift(T)Q_{\mathrm{adapt}} = Q_{\mathrm{init}} + \lambda \cdot f_{\mathrm{drift}}(T)
\qquad A = \operatorname{Softmax}\left(Q_c K_c^{\top}\right)$$ $$G(c) = \operatorname{Softmax}\left(c W_r + \mathrm{noise}\right), \qquad c = \sum_t C_t$$ The routing input c is constructed entirely from timestamps T, not from observation values Z. If there are two time series with identical timestamps but completely different numerical dynamics, they will receive identical routing weights. However, the motivation stated in the abstract is that "temporal patterns can vary substantially across samples and periods in aspects such as fluctuation dynamics". 2. The paper refers to ACSS as creating "soft segments" that capture "local variations in observation timing". However, each cursor could attend to arbitrary time points across the entire sequence because there is no local window mask to restrict the attention range.
Questions for authors
  1. Have you compared the timestamp-only router with a router that also uses observed values? This comparison would help test the claimed adaptation to fluctuation dynamics, beyond adaptation to sampling patterns.

  2. Can you compare hard versus soft segmentation and learned static versus dynamic expert weights while keeping the rest of the model unchanged? The current ablations do not clearly isolate the benefits of soft boundaries and period specific scale selection.

  3. What advantages does SoftScale offer over adapting an existing time series foundation model to the same IMTS forecasting setting, in terms of accuracy and computational cost?

  4. Could soft segmentation and dynamic scale routing be integrated into a pretrained time series foundation model (e.g., Moirai-MoE)? This would help show whether the benefits extend beyond the proposed architecture.

I would be willing to raise my score if you can satisfactorily address these concerns.

10 comments

Responses to Reviewer WMx8 (1/5)

W1/Q1: Clarification of timestamp-only routing and the role of value dynamics

We thank the reviewer for raising this point. We agree that the router receives only timestamp-derived representations at inference and therefore cannot directly identify value dynamics on its own. However, value dynamics can still help shape the routing function and influence scale weights. Specifically, the router allocates scale weights based on irregular sampling patterns, while the scale experts model observed values at different scales. During end-to-end training, value dynamics implicitly shape the learned routing function through the experts’ outputs and the forecasting loss. This allows the router to learn which scale mixtures work well for a given sampling context, taking into account how well the experts capture value dynamics during training. Our results support this.

(1) Relevance of timestamp-conditioned routing. Table 3 provides quantitative evidence that the learned scale mixtures are meaningful for forecasting. Figure 3 complements this result with a qualitative visualization of how the learned scale mixtures differ across samples and soft segments.

(2) Effect of value dynamics on router training. On PhysioNet, we perturbed only the historical values while keeping the observation timestamps and model parameters fixed. We reduced or increased the historical standard deviation by 25% while preserving the mean, or added an upward or downward linear trend using centered, range-normalized timestamps and a coefficient of 0.2 times the original standard deviation. Across 18,441 eligible histories, we compared the forecasting-loss gradients with respect to the routing logits before and after each intervention.

Value interventionGradient change ≥ 10%Gradient change ≥ 25%Gradient–Oracle agreement
Lower volatility62.3%33.7%88.1%
Higher volatility62.9%34.9%87.3%
Downward trend52.6%22.5%88.1%
Upward trend51.8%21.7%88.3%

Gradient–Oracle agreement measures how often the expert with the lowest routing-logit gradient also yields the lowest local MSE when given extra weight. With expert outputs fixed, each trial retains 80% of the original mixture and assigns the remaining 20% to one expert. Agreement was 87.3%–88.3% across the four interventions. The gradient changes show that value dynamics affect the signal used to train the router, and the high agreement shows that this signal tends to favor the scales that best serve the perturbed sequence. Through joint optimization, SoftScale learns to match scales to temporal patterns, with the router conditioned on sampling patterns and implicitly shaped by value dynamics.

(3) Directly adding observed values to the router. To test whether feeding observed values directly into the router improves forecasting, we evaluated two variants.

  • Full-history Values adds encoded observed-value features together with their mean and standard deviation, computed over each variable’s full observed history.
  • Segment-wise Values weights encoded observed-value features by the ACSS assignments and additionally includes the corresponding assignment-weighted mean and standard deviation for each soft segment.

Both variants have identical parameter counts, averaging 1.64% more parameters than SoftScale. We report MSE as mean ± sample standard deviation over five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
+ Full-history Values4.99 ± 0.064.68 ± 0.061.77 ± 0.113.22 ± 0.04
+ Segment-wise Values5.00 ± 0.054.74 ± 0.171.79 ± 0.083.22 ± 0.04

Analysis. Neither value-aware variant improved mean MSE. The gaps were larger on the clinical datasets, PhysioNet and MIMIC-III. Feeding value features directly into the router may expose it to more complex, variable, and sometimes less reliable signals, making scale allocation more sensitive to transient fluctuations. On Human Activity and USHCN, the experts may already capture recurring motion and seasonal patterns, limiting the benefit of adding value features to the router.

We will explain the rationale for this design: observation timing indicates the temporal detail supported by the data, whereas feeding observed values directly into the router may make scale weights sensitive to transient fluctuations or measurement noise.

Responses to Reviewer WMx8 (2/5)

W1/Q1 (continued). How does SoftScale match scales to temporal patterns?

SoftScale adapts to local temporal patterns through two complementary mechanisms that address sampling patterns and observed-value dynamics. The final multi-scale representation of each soft segment is jointly determined by the fusion weights and the representations extracted at each scale. The router computes the weights from the segment’s observation timing, and each scale expert extracts its representation from the observed values.

(1) Sampling patterns govern how scales are combined. SoftScale uses observation timestamps to construct sample-adaptive soft segments and assigns scale weights according to the observation timing within each segment. Consider a one-hour heart-rate record. If readings are available every few minutes, nearby observations provide more detail for modeling short-term changes at fine scales. If the same hour contains only two or three widely spaced readings, short-term changes are harder to resolve, while coarse-scale representations can summarize the overall trend using observations over a longer time span. These different sampling patterns inform how SoftScale combines the scales.

(2) Observed-value dynamics are captured in the scale-specific representations. Each expert extracts features directly from the observed values. Fine-scale representations can preserve local fluctuations, while coarse-scale representations can summarize changes over longer periods. Even when two heart-rate sequences are measured at exactly the same timestamps, a steady increase and a relatively stable trajectory can yield different features at each scale.

It is unfair to judge SoftScale’s adaptation to value dynamics solely by whether its routing weights change. Such an assessment overlooks the multi-scale information that the experts extract from the observed values. With identical fusion weights, different readings can still yield different expert representations and, consequently, different fused representations and forecasts. The fused representation learned by SoftScale reflects local sampling conditions and preserves the sample’s value dynamics across scales. In this sense, our claim that SoftScale matches scales to patterns is consistent with both the paper’s motivation and how the method constructs its multi-scale representations.

Responses to Reviewer WMx8 (3/5)

W2: Each cursor can attend to arbitrary time points without a local window mask

Thank you for pointing out that ACSS uses no local window mask. The absence of a hard local-window mask is part of the soft-segmentation design. An observation can contribute to multiple soft segments with different weights, allowing gradual transitions and information spanning segment boundaries to be retained. We examine whether each assignment develops a data-adaptive local emphasis around its corresponding cursor.

(1) Local bias from the shared continuous-time mapping. ACSS applies the same normalized continuous-time mapping h(⋅)h(\cdot) to cursors and timestamps, where h(⋅)h(\cdot) is obtained using the shared sinusoidal time embedding and MLP. We use s(qp,t)s(q_p,t) to denote the cosine similarity between cursor qpq_p and timestamp tt. Hence, s(qp,qp)=1s(q_p,q_p)=1, the maximum cosine self-similarity. This mapping is locally Lipschitz under non-degenerate normalization. Let LhL_h denote its Lipschitz constant around the cursor; then

=1-\frac{1}{2}\|h(q_p)-h(t)\|_2^2 \geq 1-\frac{L_h^2}{2}|t-q_p|^2.$$ **Thus, timestamps sufficiently close to a cursor have similarity close to the maximum, inducing a cursor-centered local bias.** Because the time embedding contains periodic components, a small number of more distant timestamps may also receive relatively high weights due to similar representations. **(2) Locality audit.** To verify whether the ACSS assignments exhibit locality, we conducted the following experiment using five independently trained checkpoints on each of the four datasets. - **Set construction.** For each soft segment, we identified the **20% of valid observations with the highest attention weights** and the **20% closest to the cursor** in time, rounding both counts up. - **Top-20% proximity mass.** This metric measures the fraction of total attention weight in the high-attention set that comes from observations also belonging to the nearest-to-cursor set. - **Matched random baseline.** We keep the top-attention group and its weights fixed, replace the $k$ observations closest to the cursor with $k$ observations drawn uniformly at random from the $n$ valid observations, and compute the same attention-weight share. Each observation has a $k/n$ chance of being drawn, so the expected share is $k/n$. For example, with seven valid observations, rounding up the top 20% gives $k=2$, so the baseline is $2/7\approx28.6\%$. | Dataset | Top-20% mass in nearest 20% | Matched random baseline | Holm-adjusted $p$ | |---|---:|---:|---:| | PhysioNet | $70.48\pm4.20\%$ | 31.22% | 0.00043 | | Activity | $41.13\pm6.90\%$ | 22.66% | 0.0157 | | USHCN | $47.93\pm2.28\%$ | 21.18% | 0.00019 | | MIMIC | $79.08\pm2.88\%$ | 40.35% | 0.00012 | | Dataset mean | **59.66%** | **28.85%** | — | **The proportion was significantly above** the matched random baseline on all four datasets after Holm correction. This indicates that high-weight observations tend to concentrate near their corresponding cursors. The gaps above the random baseline were smaller on Activity and USHCN. One possible explanation is that valid observations are denser on these datasets, leaving more candidates at similar distances from the cursor. A top-20% cutoff can place temporally close observations on opposite sides of the threshold, reducing the measured overlap. Activity also contains smoothly evolving inertial-sensor trajectories, while USHCN exhibits seasonal cycles, so more distant observations from similar phases of the seasonal cycle may provide context. These two datasets may therefore require less sharply peaked local attention than the highly sparse clinical datasets. **Overall, ACSS emphasizes local observations while allowing the range of attention to vary across segments.** Observations near the cursor provide local temporal context for segment-level scale adaptation. When nearby observations are sparse or relevant patterns span longer periods, soft assignment allows more distant observations to contribute.
Reviewer_jq5aOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SoftScale for irregular multivariate time-series forecasting. The method first uses learnable temporal cursors to softly partition irregular observations into several latent temporal segments. It then processes each segment with multiple experts operating at predefined temporal scales and uses a noisy mixture-of-experts router to aggregate their outputs. Cross-variable attention is further introduced within each scale. Experiments on four irregular time-series benchmarks report consistently competitive results against a broad collection of baselines. The paper is technically coherent and empirically solid, but its claimed pattern-adaptive scale selection is not fully supported by the actual routing mechanism, which is driven primarily by timestamps rather than observed temporal patterns.

Critical strengths
  1. The paper addresses a relevant limitation of fixed-scale irregular time-series models. The observation that different temporal regions may require different aggregation scales is reasonable, particularly when sampling density and temporal dynamics vary within the same sequence.
  2. The architecture is coherent and end-to-end trainable. Soft temporal segmentation avoids hard boundary decisions, while the combination of temporal aggregation and cross-variable modeling is appropriate for irregular multivariate observations.
Critical weaknesses
  1. The routing mechanism does not directly observe the value patterns whose scales it is claimed to identify. Both the soft-segment assignments and the expert-routing weights are derived from timestamp embeddings. Consequently, two samples with identical observation times but substantially different amplitudes, volatility, trends, or local frequencies receive the same scale-routing pattern. This contradicts the central claim that SoftScale matches scales to sample-specific fluctuation dynamics rather than primarily to sampling configurations.
  2. The methodological novelty is limited. SoftScale mainly combines learnable time-query attention, predefined multi-scale temporal aggregation, per-scale cross-variable attention, and conventional noisy MoE routing. These are established components, and the paper does not establish a sufficiently distinctive learning principle beyond their integration.
  3. The claimed “period-specific” segmentation is weakly defined. The soft segments are attention distributions over observations rather than temporally ordered, non-overlapping intervals. There is no clear mechanism ensuring that different cursors represent distinct chronological periods instead of overlapping or redundant collections of observations, making the interpretation of each segment as a temporal period questionable.
  4. The mechanism-level evidence is considerably weaker than the forecasting results. Removing multi-scale modeling or perturbing the routing produces relatively modest degradation on several datasets. The main qualitative analysis is also restricted to a small case study, so the reported accuracy improvements do not convincingly demonstrate that the model succeeds specifically by matching temporal scales to local patterns.
  5. The method remains dependent on manually predefined structural choices. The candidate scale pool and number of temporal cursors are dataset-level hyperparameters, while all scale experts are evaluated through dense routing. The method therefore selects among manually supplied scales rather than learning the relevant temporal resolutions, and it does not obtain the computational sparsity commonly associated with MoE architectures.
Questions for authors

See Weaknesses.

9 comments

Responses to Reviewer jq5a (1/5)

W1: Timestamp-only routing does not directly observe the value patterns it is claimed to identify

We thank the reviewer for raising this point. We agree that the router receives only timestamp-derived representations at inference and therefore cannot directly identify value dynamics on its own. However, value dynamics can still help shape the routing function and influence scale weights. Specifically, the router allocates scale weights based on irregular sampling patterns, while the scale experts model observed values at different scales. During end-to-end training, value dynamics implicitly shape the learned routing function through the experts’ outputs and the forecasting loss. This allows the router to learn which scale mixtures work well for a given sampling context, taking into account how well the experts capture value dynamics during training. Our results support this.

(1) Relevance of timestamp-conditioned routing. Table 3 provides quantitative evidence that the learned scale mixtures are meaningful for forecasting. Figure 3 complements this result with a qualitative visualization of how the learned scale mixtures differ across samples and soft segments.

(2) Effect of value dynamics on router training. On PhysioNet, we perturbed only the historical values while keeping the observation timestamps and model parameters fixed. We reduced or increased the historical standard deviation by 25% while preserving the mean, or added an upward or downward linear trend using centered, range-normalized timestamps and a coefficient of 0.2 times the original standard deviation. Across 18,441 eligible histories, we compared the forecasting-loss gradients with respect to the routing logits before and after each intervention.

Value interventionGradient change ≥ 10%Gradient change ≥ 25%Gradient–Oracle agreement
Lower volatility62.3%33.7%88.1%
Higher volatility62.9%34.9%87.3%
Downward trend52.6%22.5%88.1%
Upward trend51.8%21.7%88.3%

Gradient–Oracle agreement measures how often the expert with the lowest routing-logit gradient also yields the lowest local MSE when given extra weight. With expert outputs fixed, each trial retains 80% of the original mixture and assigns the remaining 20% to one expert. Agreement was 87.3%–88.3% across the four interventions. The gradient changes show that value dynamics affect the signal used to train the router, and the high agreement shows that this signal tends to favor the scales that best serve the perturbed sequence. Through joint optimization, SoftScale learns to match scales to temporal patterns, with the router conditioned on sampling patterns and implicitly shaped by value dynamics. p

(3) Directly adding observed values to the router. To test whether feeding observed values directly into the router improves forecasting, we evaluated two variants.

  • Full-history Values adds encoded observed-value features together with their mean and standard deviation, computed over each variable’s full observed history.
  • Segment-wise Values weights encoded observed-value features by the ACSS assignments and additionally includes the corresponding assignment-weighted mean and standard deviation for each soft segment.

Both variants have identical parameter counts, averaging 1.64% more parameters than SoftScale. We report MSE as mean ± sample standard deviation over five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
+ Full-history Values4.99 ± 0.064.68 ± 0.061.77 ± 0.113.22 ± 0.04
+ Segment-wise Values5.00 ± 0.054.74 ± 0.171.79 ± 0.083.22 ± 0.04

Analysis. Neither value-aware variant improved mean MSE. The gaps were larger on the clinical datasets, PhysioNet and MIMIC-III. Feeding value features directly into the router may expose it to more complex, variable, and sometimes less reliable signals, making scale allocation more sensitive to transient fluctuations. On Human Activity and USHCN, the experts may already capture recurring motion and seasonal patterns, limiting the benefit of adding value features to the router.

We will explain the rationale for this design: observation timing indicates the temporal detail supported by the data, whereas feeding observed values directly into the router may make scale weights sensitive to transient fluctuations or measurement noise.

Responses to Reviewer jq5a (2/5)

W1 (continued). How does SoftScale match scales to temporal patterns?

SoftScale adapts to local temporal patterns through two complementary mechanisms that address sampling patterns and observed-value dynamics. The final multi-scale representation of each soft segment is jointly determined by the fusion weights and the representations extracted at each scale. The router computes the weights from the segment’s observation timing, and each scale expert extracts its representation from the observed values.

(1) Sampling patterns govern how scales are combined. SoftScale uses observation timestamps to construct sample-adaptive soft segments and assigns scale weights according to the observation timing within each segment. Consider a one-hour heart-rate record. If readings are available every few minutes, nearby observations provide more detail for modeling short-term changes at fine scales. If the same hour contains only two or three widely spaced readings, short-term changes are harder to resolve, while coarse-scale representations can summarize the overall trend using observations over a longer time span. These different sampling patterns inform how SoftScale combines the scales.

(2) Observed-value dynamics are captured in the scale-specific representations. Each expert extracts features directly from the observed values. Fine-scale representations can preserve local fluctuations, while coarse-scale representations can summarize changes over longer periods. Even when two heart-rate sequences are measured at exactly the same timestamps, a steady increase and a relatively stable trajectory can yield different features at each scale.

It is unfair to judge SoftScale’s adaptation to value dynamics solely by whether its routing weights change. Such an assessment overlooks the multi-scale information that the experts extract from the observed values. With identical fusion weights, different readings can still yield different expert representations and, consequently, different fused representations and forecasts. The fused representation learned by SoftScale reflects local sampling conditions and preserves the sample’s value dynamics across scales. In this sense, our claim that SoftScale matches scales to patterns is consistent with both the paper’s motivation and how the method constructs its multi-scale representations.

Responses to Reviewer jq5a (3/5)

W2: Limited methodological novelty beyond integrating established components

In IMTS, the dynamics of observed values and sampling density can vary across samples and over time within a sample. The core idea of SoftScale is to adapt scales at the level of sample-adaptive soft segments and use the temporal context of each segment to guide both multiscale feature extraction and scale fusion. This design addresses two types of variation:

  1. Scale adaptation across samples. SoftScale adjusts its time cursors based on each sample’s original observation timestamps and uses time-query attention to construct soft segments. The temporal context that guides scale adaptation can thus reflect each sample’s irregular temporal patterns.
  2. Scale adaptation within a sample through soft segmentation. SoftScale learns a separate scale mixture for each soft segment within a sample. When temporal patterns change gradually, observations in transition regions can contribute to multiple segments with different weights. This reduces the separation of observations with similar patterns on opposite sides of a hard segment boundary.

Each mechanism in SoftScale is tailored to this problem. To adapt to irregular temporal patterns across samples, we adjust the time cursors for each sample based on its original observation timestamps and use time-query attention to construct soft segments. This allows observations in transition regions to contribute to multiple segments with different weights. Each scale expert extracts features from the original observed values at its own scale and models dependencies across variables at that scale. The MoE router uses the temporal representation of each soft segment to learn scale fusion weights for each variable in that segment. During joint training, the dynamics of the observed values modeled by the experts also influence the learned weights through the forecasting loss. These task-specific designs enable SoftScale to match scales to patterns through the way the mechanisms work together. Simply integrating the components would not give the model this capability without making soft-segment representations the common basis for both feature extraction and scale fusion.

W5: Dependence on predefined structural choices and dense expert routing

(1) Candidate scale bases and input-driven mixed scales. We agree that the candidate scale pool and the number of temporal cursors are predefined. SoftScale's innovation lies in learning input-driven mixed-scale representations for different samples, variables, and soft segments over a finite set of interpretable scale bases, thereby enabling adaptive multi-scale modeling.

(2) Motivation for dense routing. The small set of experts in SoftScale models different temporal scales to integrate complementary scale information. The same time series, and even the same soft segment, may simultaneously contain local fluctuations, medium-term changes, and long-term trends. In sparse, asynchronous IMTS, limited observations may introduce uncertainty into local scale selection. Dense routing preserves all scale-specific views and adjusts their contributions through continuous weights, allowing more flexible scale combinations. This design evaluates all scale experts and therefore does not provide computational sparsity.

(3) Top-2 fusion comparison. To examine whether explicitly discarding some scales benefits forecasting, we added a Top-2 fusion ablation on all four datasets. It selects only the two experts with the highest routing weights and renormalizes their weights before fusion. The table reports MSE as mean ± sample standard deviation across five seeds.

ModelUSHCN (×10⁻¹)PhysioNet (×10⁻³)MIMIC-III (×10⁻²)Human Activity (×10⁻³)
SoftScale4.98 ± 0.074.57 ± 0.071.67 ± 0.033.19 ± 0.02
Top-2 fusion5.06 ± 0.074.92 ± 0.101.74 ± 0.023.23 ± 0.05

Results and analysis. Dense routing achieves lower mean MSE on all four datasets, with larger relative gaps on PhysioNet and MIMIC-III. Clinical observations are often sparse and asynchronous, and abrupt local changes can coexist with slower variations. Discarding a scale may therefore remove complementary temporal information. Motion continuity and repeated actions in Human Activity, together with the seasonality of weather variables in USHCN, may yield more regular dominant patterns and reduce the impact of discarding a scale. The overall results support preserving complementary scale-specific views and using continuous weights to form input-driven mixed-scale representations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.