- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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 (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 intervention | 95% 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 -tests across five training seeds ().
(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.
| Dataset | Ablation | MSE (mean ± SD) | MSE wins | MAE (mean ± SD) | MAE wins | Parameters | Training epochs (mean ± SD) |
|---|---|---|---|---|---|---|---|
| MIMIC-III | w/o ACSS | 5/5 | 5/5 | 708,389 / 707,555 | 41.2 ± 5.5 / 66.6 ± 24.3 | ||
| MIMIC-III | w/o MoE | 5/5 | 4/5 | 708,389 / 708,389 | 41.2 ± 5.5 / 46.0 ± 9.9 | ||
| PhysioNet | w/o ACSS | 5/5 | 5/5 | 1,070,208 / 1,069,484 | 47.6 ± 7.3 / 57.8 ± 22.6 | ||
| PhysioNet | w/o MoE | 5/5 | 3/5 | 1,070,208 / 1,070,208 | 47.6 ± 7.3 / 53.8 ± 10.4 | ||
| Human Activity | w/o ACSS | 5/5 | 3/5 | 222,857 / 222,511 | 88.8 ± 6.9 / 112.6 ± 31.3 | ||
| Human Activity | w/o MoE | 5/5 | 4/5 | 222,857 / 222,857 | 88.8 ± 6.9 / 78.2 ± 7.4 | ||
| USHCN | w/o ACSS | 3/5 | 5/5 | 157,748 / 157,416 | 81.2 ± 4.9 / 101.8 ± 16.3 | ||
| USHCN | w/o MoE | 4/5 | 3/5 | 157,748 / 157,748 | 81.2 ± 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.
| Model | USHCN (×10⁻¹) | PhysioNet (×10⁻³) | MIMIC-III (×10⁻²) | Human Activity (×10⁻³) |
|---|---|---|---|---|
| SoftScale | 4.98 ± 0.07 | 4.57 ± 0.07 | 1.67 ± 0.03 | 3.19 ± 0.02 |
| Global Static | 5.10 ± 0.04 | 4.76 ± 0.19 | 1.78 ± 0.10 | 3.27 ± 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.
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.
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.
Both variants have identical parameter counts, averaging 1.64% more parameters than SoftScale. We report MSE as mean ± sample standard deviation over five seeds.
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.