acceptodds
Under review as a conference paper at ICLR 2027

SEQNet: Support-Aware Irregular Forecasting with Event Memory and Query Routing

Abstract

Irregular observations are common in healthcare, activity monitoring, and climate analysis. Forecasting from such data requires compressing histories with varying lengths and irregular time intervals into fixed-length representations. However, although historical compression can capture overall temporal trends, it may overlook a small number of locally predictive observations, such as abnormal values and abrupt changes, making these critical local signals difficult to fully exploit. Moreover, when the target variable has insufficient historical observations and information from other variables is needed, different forecasting targets may require different historical evidence and information sources, making fixed information-selection strategies difficult to adapt to changing observation conditions.We propose SEQNet, a forecasting model that preserves both global historical trends and prediction-relevant local evidence. Specifically, SEQNet summarizes historical observations to learn variable-level temporal patterns while constructing a candidate set of local events and dynamically retrieving observations that are relevant to the current forecasting query. When the target variable is insufficiently observed, SEQNet further adapts to its current observation support by assigning different weights to the target variable's own history, information from other variables, and general patterns learned from data.Experiments on PhysioNet, Activity, USHCN, and MIMIC-III show that SEQNet achieves the lowest MSE among the compared methods on all four datasets. On PhysioNet, SEQNet reduces MSE by 24.49% relative to Hi-Patch, the strongest baseline.

Primary area:
learning on time series and dynamical systems
Keywords:
irregular time series, multivariate time series forecasting, continuous volatility indexing, adaptive pruning, robust statistics, Transformer, missing data, attention collapse

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_MKu8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SEQNet for irregular multivariate time-series forecasting. It combines compressed temporal memory with explicitly retained local events and uses query-conditioned, support-aware routing to balance self-history, cross-variable information, and learned priors. Experiments on four datasets show the lowest MSE among the compared methods, with additional ablations on event retention and routing.

Critical strengths
  1. The combination of compressed global memory, retained local events, and support-aware routing is well motivated and directly addresses important characteristics of irregular time series.

  2. The empirical results are strong, with the best MSE on all four datasets and results reported over five random seeds.

Critical weaknesses
  1. The observation-deletion robustness experiment evaluates only SEQNet. This makes it difficult to determine whether its support-aware design is actually more robust than competing methods. The 416% average MSE increase on Activity makes this comparison particularly important.

  2. Several training mechanisms are introduced together, including paired-view training, constrained family weighting, and auxiliary event supervision, but their individual contributions are not sufficiently isolated.

Questions for authors
  1. How does SEQNet compare with strong baselines such as Hi-Patch or t-PatchGNN under the same observation-deletion environments, particularly on Activity?

  2. What are the individual contributions of paired-view training, constrained family weighting, and event supervision? Could the authors provide ablations for these training components?

4 comments

Response to Reviewer MKu8


Thank you for recognizing the motivation behind SEQNet's design, which combines a summary of the history with direct access to selected observations and uses the available history to decide which sources to rely on. We also appreciate your noting that SEQNet achieved the lowest MSE among the compared methods on all four benchmarks, with results reported over five seeds. To address your questions, we compare SEQNet with strong baselines under observation deletion on Activity and PhysioNet, and test each training component with the architecture held fixed.


W1 & Q1. Comparison with strong baselines under the same observation-deletion environments

Activity. On the revised subject-disjoint split, we evaluated SEQNet, Hi-Patch, and t-PatchGNN on the same test samples using sample-micro MSE (Appendix E.3, Eq. (17)). We used the five observation-deletion families and four retention rates from §§3.5 and 4.3, yielding 20 test environments. Each result is from a single run (seed 1).

ModelTrainingClean MSEMean MSE over 20 environmentsWorst-environment MSEMean MSE increase over clean
t-PatchGNNClean-only0.0025110.0780310.2006213007.98%
Hi-PatchClean-only0.0023470.0095130.028471305.30%
SEQNetPaired deletion0.0022320.0032100.00818343.81%

On held-out subjects, SEQNet achieves the lowest clean, mean, and worst-environment MSE. Its mean and worst-environment MSE are 66.26% and 71.26% lower than Hi-Patch's and 95.89% and 95.92% lower than t-PatchGNN's, respectively.

The 416.48% increase in Table 3 of the submission was measured on the original window-level split. On the subject-disjoint split, SEQNet's mean MSE increases by 43.81% over its clean MSE. The worst environment remains variable-subset deletion at a retention rate of 0.2, with an MSE of 0.008183.

PhysioNet. We also trained both baselines with paired deletion and evaluated them using the same five deletion families, four retention rates, and sample-micro MSE.

ModelTrainingClean MSEMean MSE over 20 environmentsWorst-environment MSE
t-PatchGNNClean-only0.0041700.0059840.009235
t-PatchGNNPaired deletion0.0041320.0050630.007637
Hi-PatchClean-only0.0041620.0067340.012979
Hi-PatchPaired deletion0.0041690.0050680.007659
SEQNetPaired deletion0.0036700.0045030.006743

Paired-deletion training reduces mean and worst-environment MSE by 15.39% and 17.30% for t-PatchGNN and by 24.74% and 40.99% for Hi-Patch. With paired-deletion training applied to both baselines, SEQNet's mean and worst-environment MSE are still 11.06% and 11.71% lower than t-PatchGNN's and 11.15% and 11.96% lower than Hi-Patch's, respectively. We will add both comparisons to §4.3 and update the Activity results in Table 3.

To make it easier to reproduce our results, we provide the code in an anonymous repository.


W2 & Q2. Individual contributions of the three training components

We held the forecasting architecture fixed and ablated paired views, constrained family weighting, and auxiliary event supervision on PhysioNet. Paired-view training predicts the same targets from the original and deleted histories (§3.5, Eq. (5)). Both weighting configurations sample deletion families uniformly. Constrained weighting adjusts each family's contribution to the deleted-view forecasting loss (Eq. (6)). Auxiliary event supervision uses binary cross-entropy for lower- and upper-tail labels defined by the 5th and 95th percentiles of the training data (§3.6, Eq. (7)).

ConfigurationPaired viewsDeletion-family weightingEvent supervisionClean test MSE
D0Off—Off0.003801
D1OnUniformOff0.003748
D2OnConstrainedOff0.003736
D3OnUniformOn0.003706
D4OnConstrainedOn0.003689
D5Off—On0.003787
  • Paired views. Adding paired views with uniform weighting reduces MSE by 1.39% without event supervision (D0→D1) and 2.14% with it (D5→D3).
  • Constrained weighting. With paired views enabled, constrained weighting further reduces MSE by 0.32% (D1→D2) and 0.46% (D3→D4).
  • Auxiliary event supervision. Adding event supervision reduces MSE by 0.37% (D0→D5), 1.12% (D1→D3), and 1.26% (D2→D4) in the three matched comparisons.

Due to the character limit, please see the remaining questions in the following comment.


Continued from previous comment


Paired views produce the largest individual reduction in this single-run ablation. With all three components enabled, D4 achieves the lowest MSE, 2.95% below D0. We will add the full ablation to §4.4 and Appendix F and identify the loss terms for each component in §§3.5–3.6.


Finally, thank you once again for the patient and thorough reading of our paper and recognizing its strengths. Your thoughtful questions and suggestions helped us improve the manuscript. We would be happy to continue the discussion if you have any further questions!


I was wondering if our response has addressed your concern... If you feel that your concerns have all been addressed, we would be very grateful if you could consider giving it a higher score... But regardless, we sincerely thank you for carefully reading our paper and for your excellent suggestions, which have helped us improve it a lot!^^

AuthorsOpenReview ↗

Sorry for the quick follow-up...T T

As the rebuttal deadline is approaching, I’m sorry to bother you again... but we hope this message does not bring you any pressure at all...!!

We are very grateful for your initial recognition of this paper, and we feel truly happy and lucky that it was reviewed by you. We just wanted to confirm whether your concerns have been addressed. We really hope our response can make our paper more convincing.

Finally, we understand you may be very busy in real life. But regardless, hoping all your papers get accepted!~

Reviewer_hUB8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SEQNet combines compressed temporal memory, retained local events, and support-aware query routing for irregular multivariate forecasting, achieving promising results across four datasets, although the comparative evaluation and robustness evidence require clarification.

Critical strengths
  1. The combination of temporal summaries, retained observations, and support-aware routing provides a well-motivated approach to forecasting from sparse or stale histories.
  2. SEQNet reports the lowest MSE across four datasets over five seeds, including substantial improvements of 24.49% on PhysioNet and 15.18% on Activity. 3.The matched-budget event-selection experiments and source-routing ablations provide evidence that event identity and adaptive source fusion contribute beyond simply adding memory tokens.
Critical weaknesses
  1. The paper does not establish whether all baselines use identical preprocessing, splits, and metric reductions, while several recent methods discussed in related work are absent from the comparison.
  2. The robustness evaluation includes no competing methods and shows a 416.48% average MSE increase on Activity, leaving the claimed benefits under observation deletion insufficiently supported.
  3. The ablations do not independently isolate paired-view training, constrained family weighting, and auxiliary event supervision, limiting attribution of the gains to the proposed architecture.
Questions for authors
  1. How does SEQNet’s observation-deletion training differ from TameR’s randomized sampling (Xu et al., ICML 2026), and does retaining recent, high-deviation, and high-change observations increase sensitivity to corrupted inputs or recent-data bias?
  2. Were all baselines retrained using identical splits, normalization, target definitions, and metric reductions, and how does SEQNet compare with omitted methods such as ASTGI, APN, and HyperIMTS?
  3. Do the Activity window-level and MIMIC-III instance-level splits prevent overlapping observations or the same subject or patient from appearing across partitions?
  4. How does SEQNet compare with strong baselines under the same deletion environments and matched augmentation, particularly in the Activity settings responsible for the large degradation?
  5. What are the absolute MSEs and uncertainties for the progressive configurations, including the exact clean reference, and for independent ablations of paired-view training, family weighting, and event supervision?

Reference: Xu, L., et al. (2026). Taming the Recent-Data Bias: Towards Robust Time Series Forecasting with Global Context. ICML.

6 comments
AuthorsOpenReview ↗

Response to Reviewer hUB8

Response to Reviewer hUB8

Thank you for recognizing why SEQNet combines temporal summaries, retained observations, and support-aware routing when history is sparse or outdated. We also appreciate your noting its low forecasting error across the four benchmarks and the evidence from the equal-budget event selection and routing ablations. Below we report the new baseline comparisons, the deletion results on held-out Activity subjects, and the training ablations.

W1 & Q2. Baseline comparisons and evaluation protocols

Table 1 and Appendix E.1 describe dataset construction, splits, and forecast horizons. Equations (17)–(19) in Appendix E.3 define the MSE reductions. The table below summarizes the evaluation protocols, including the revised subject-disjoint Activity split.

DatasetSplit unitTrain / validation / test samplesHistory → forecast horizonNormalizationMSE reduction
PhysioNetPatient ID, split before eligibility filtering7,187 / 2,396 / 2,3920–24 h → 24–48 hTraining-only per-channel min–maxAverage within each sample, then across samples
MIMIC-IIIForecasting instance14,073 / 4,692 / 4,6920–24 h → 24–48 hTraining-only min–maxAverage over all valid target entries
ActivitySubject, 15 / 5 / 5 subjects6,598 / 2,001 / 2,1403 s → following 1 sPer-channel min–max fitted on training subjectsAverage within each sample, then across samples
USHCNStation, split before window construction15,715 / 5,224 / 5,25924 months → following monthProcessed raw valuesAverage within each channel, then across channels

We added APN, ASTGI, and HyperIMTS. The Activity results use the subject-disjoint split.

MethodPhysioNetMIMIC-IIIActivity (subject-disjoint)USHCN
SEQNet0.00367 ± 0.000050.0920 ± 0.00060.0022320.495 ± 0.005
APN0.0040600.1014210.0026810.558340
ASTGI0.0037820.0987970.0023320.527501
HyperIMTS0.0049500.2033000.0025300.549600
SEQNet MSE reduction vs. best added baseline2.96%6.88%4.29%6.16%

SEQNet leads the three recent baselines on all four datasets. Compared with the strongest added baseline on each dataset, its MSE is 2.96% lower on PhysioNet, 6.88% lower on MIMIC-III, 4.29% lower on Activity, and 6.16% lower on USHCN. The Activity advantage holds on subjects the models never saw during training. We will add the results to §4.2 and document the evaluation settings in Appendix E.

To make it easier to reproduce our results, we provide the code in an anonymous repository.


W2 & Q4. Comparisons under the same observation-deletion environments

Sections 3.5 and 4.3 define five deletion families: random observations, variable subsets, contiguous time blocks, early observations, and recent observations. Each family is evaluated at four retention rates (0.8, 0.6, 0.4, and 0.2), giving 20 test environments. Table 3 reports the following SEQNet results.

DatasetClean MSEMean MSE over 20 environmentsWorst-environment MSEMean MSE increase over clean
PhysioNet0.0036700.0045030.00674322.70%
USHCN0.4950000.5259790.5813156.26%
Activity (original window-level split)0.0021800.0112590.069881416.48%

On PhysioNet, we evaluated t-PatchGNN and Hi-Patch using the same deletion families, retention rates, and sample-micro MSE. For each baseline, we compared clean-only training with paired-deletion training. The final column measures the increase relative to each configuration's own clean MSE.

ModelTrainingClean MSEMean MSE over 20 environmentsWorst-environment MSEMean MSE increase over clean
t-PatchGNNClean only0.0041700.0059840.00923543.50%
t-PatchGNNPaired deletion0.0041320.0050630.00763722.53%
Hi-PatchClean only0.0041620.0067340.01297961.80%
Hi-PatchPaired deletion0.0041690.0050680.00765921.56%
SEQNetPaired deletion0.0036700.0045030.00674322.70%

Paired-deletion training reduces mean and worst-environment MSE by 15.39% and 17.30% for t-PatchGNN and by 24.74% and 40.99% for Hi-Patch. Among the three paired-training models, mean MSE increases by 21.56%–22.70% relative to each model's clean MSE. In absolute terms, SEQNet's mean and worst-environment MSE are 11.06% and 11.71% lower than paired t-PatchGNN and 11.15% and 11.96% lower than paired Hi-Patch.


Due to the character limit, please see the remaining questions in the following comment.


Continued from previous comment


Activity on held-out subjects. The 416.48% increase in Table 3 came from the original window-level split, where deleting whole variables was particularly damaging. At a retention rate of 0.2, the MSE was 0.069881 for variable-subset deletion, compared with 0.003017 for random point deletion and 0.002836 for time-block deletion. We have now evaluated SEQNet and both baselines on the same subject-disjoint test samples across all 20 deletion environments, using sample-micro MSE. Each result is from a single run (seed 1).

ModelTrainingClean MSEMean MSE over 20 environmentsWorst-environment MSEMean MSE increase over clean
t-PatchGNNClean only0.0025110.0780310.2006213007.98%
Hi-PatchClean only0.0023470.0095130.028471305.30%
SEQNetPaired deletion0.0022320.0032100.00818343.81%

On this split, SEQNet's mean MSE rises by 43.81% from clean, compared with 416.48% on the original split. Its hardest condition is still variable-subset deletion at a retention rate of 0.2, with an MSE of 0.008183. SEQNet has the lowest clean, mean, and worst-environment MSE. Compared with Hi-Patch, the stronger Activity baseline in this table, its mean and worst-environment MSE are 66.26% and 71.26% lower. The PhysioNet comparison above shows that SEQNet continues to lead when both baselines also use paired-deletion training. We will add both comparisons to §4.3 and update the Activity results in Table 3.


W3 & Q5. Progressive construction and independent training ablations

Section 4.4, Figure 3(a), and Table 7 report the progressive construction experiment on PhysioNet. Each reduction is relative to the same clean reference.

Progressive configurationMSE reduction relative to the clean reference
Clean reference0.00%
Soft Temporal Memory2.19%
Add Retained Events3.05%
Add query-conditioned retrieval3.70%
Full SEQNet4.35%

Table 2 reports 0.00367 ± 0.00005 for full SEQNet on PhysioNet over five seeds.

The following training ablation holds the SEQNet architecture fixed. Paired views predict the same targets from the original and deleted histories (§3.5, Eq. (5)). Both weighting configurations sample deletion families uniformly. Constrained weighting adjusts family weights in the deleted-view forecasting loss (Eq. (6)). Event supervision uses binary cross-entropy for lower- and upper-tail labels defined by the training-set 5th and 95th percentiles (§3.6, Eq. (7)).

ConfigurationPaired viewsDeletion-family weightingEvent supervisionClean test MSE
D0Off—Off0.003801
D1OnUniformOff0.003748
D2OnConstrainedOff0.003736
D3OnUniformOn0.003706
D4OnConstrainedOn0.003689
D5Off—On0.003787

Adding paired views with uniform weighting reduces MSE by 1.39% without event supervision (D0→D1) and 2.14% with it (D5→D3). With paired views enabled, constrained weighting reduces MSE by 0.32% (D1→D2) and 0.46% (D3→D4). Adding event supervision reduces MSE in all three matched comparisons, by 0.37% (D0→D5), 1.12% (D1→D3), and 1.26% (D2→D4). D4 has the lowest MSE in this single-run ablation, 2.95% below D0. We will add the absolute MSEs to §4.4 and Appendix F and identify the loss terms for each component in §§3.5–3.6.

Does event selection also help other models? We added the same event selector to t-PatchGNN and Hi-Patch on USHCN. All configurations use paired-deletion training. The Random and Mixed configurations use the same event branch and retain the same number of observations. Only the choice of observations changes.

USHCN backbonePaired, no eventsPaired, random eventsPaired, Mixed events
t-PatchGNN0.6298720.6215520.611411
Hi-Patch0.6786030.6524600.625265

Mixed selection lowers MSE by 1.63% for t-PatchGNN and 4.17% for Hi-Patch compared with keeping the same number of random observations. This shows that choosing which observations to keep helps both backbones, even with the same event budget and training setup. We will report the transfer results in Appendix F.


Due to the character limit, please see the remaining questions in the following comment.


Continued from previous comment


Q1. Comparison with TameR and sensitivity to recent-value corruption

TameR uses basis-aligned randomized sampling, periodicity extraction, and two-stage training to reduce reliance on individual recent observations. SEQNet's paired training varies observation support across five deletion families while holding the forecast queries and targets fixed. At inference, SEQNet retrieves evidence from compressed history and individually retained events, then uses observation support to weight the target variable's own history, other variables' histories, and a learned prior (§§3.3–3.5). TameR, ICML 2026

We added clean forecasting comparisons with TameR on all four datasets. Each TameR result is from one run (seed 1). Activity uses the subject-disjoint split. The MIMIC-III SEQNet value in this table is from the direct comparison.

DatasetTameR MSESEQNet MSESEQNet MSE reduction
PhysioNet0.0041590.00367011.76%
Activity (subject-disjoint)0.0028010.00223220.31%
USHCN0.6987730.49500029.16%
MIMIC-III0.1713410.09843542.55%

SEQNet has lower clean MSE than TameR on all four datasets. We will add this comparison to §4.2.

Value corruption can increase an observation's deviation or local-change score and make it more likely to be retained. Table 10 compares the following event-selection rules on USHCN.

Event-selection ruleUSHCN test MSE
Recency only0.505680
Deviation only0.505400
Local change only0.500841
Random events0.511604
Full selection rule0.495000

The full selection rule achieves lower MSE than recency-only selection. Recent-observation deletion measures the impact of missing recent information. Quantifying sensitivity to recent-value corruption requires applying perturbations of matched count and magnitude at recent, early, and random positions while holding the model fixed. We will clarify the distinction from TameR in related work and specify the perturbations covered by the current evaluation in §4.3.


Q3. Activity and MIMIC-III split boundaries

Activity. The original window-level split could place overlapping windows in different partitions. The revised protocol assigns the 25 subjects to train, validation, and test sets (15/5/5) before generating windows within each partition. There is zero overlap in subjects and raw observations across partitions. Normalization statistics are fitted on the training subjects.

MethodSubject-disjoint Activity test MSE
SEQNet0.002232
ASTGI0.002332
Hi-Patch0.002347
t-PatchGNN0.002511
HyperIMTS0.002530
APN0.002681
TameR0.002801

SEQNet retains its lead over all six baselines on held-out subjects with zero cross-partition observation overlap. Its MSE is 4.29% lower than ASTGI, the strongest baseline in this comparison. We will update Tables 1 and 2 and Appendix E.1 and provide the subject lists and split code.

MIMIC-III. We use a fixed 60/20/20 split of 14,073/4,692/4,692 forecasting instances. Each of the 23,457 patients contributes one forecasting instance and appears in exactly one partition. The training, validation, and test sets share no patients or raw observations. Within each instance, observations from 0–24 h form the history and those from 24–48 h serve as prediction targets. The two windows do not overlap.


Thank you again for your constructive and valuable questions. We hope the additional experiments and clarifications address your concerns. If you have any other questions or suggestions, we look forward to continuing the conversation!


Reviewer_7LU8OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes SEQNet for irregular multivariate time-series forecasting with events. Its main contribution is a dual-memory architecture that combines compressed temporal summaries with retained observations selected by recency, deviation, and local change. A query-conditioned decoder retrieves relevant evidence and uses observation-support descriptors to route among target-history, cross-variable, and learned-prior representations. Training combines original and observation-deleted histories with constrained deletion-family reweighting and auxiliary tail-event supervision.

Critical strengths
  1. The paper addresses two worthwhile problems: historical compression may weaken important local observations, and selecting information sources under sparse observations is challenging because observation counts, sampling intervals, and recency vary substantially across variables.

  2. The dual-memory architecture aligns well with these motivations. Soft memory aggregates historical patterns, while the event branch preserves local observations for direct retrieval. It keeps the design clear.

Critical weaknesses

1.The contribution beyond existing methods is unclear. The method combines existing components in a targeted way, but its contribution beyond prior aggregation, query-based decoding, and gated fusion approaches remains unclear. Although it combines soft temporal aggregation, event selection, query attention, and gated fusion, the paper does not sufficiently explain what is fundamentally new relative to related methods or why the proposed design is necessary.

2.Several recent baselines are missing. The related work discusses APN, ASTGI, and HyperIMTS, but Table 2 does not include them. In particular, APN and ASTGI, published at AAAI 2026 and ICLR 2026, respectively, are directly relevant to adaptive aggregation and query-based prediction. It is also unclear whether all baselines use the same data splits, preprocessing, and evaluation metrics.

3.The contribution of event retention needs further clarification. The existing ablations show that retained events help SEQNet, but do not establish whether the same strategy would provide similar gains for existing models. Please compare against a strong irregular forecasting baseline, such as Hi-Patch or t-PatchGNN, augmented with the same event selection and retention strategy under comparable model capacity and training settings. This would help distinguish the benefit of event retention from that of the overall architecture.

  1. Training strategies are insufficiently ablated,we do not know which part is actually effective. The paper includes event-selection and routing ablations, but does not isolate the effects of paired-view training, constrained family weighting, or event auxiliary supervision. Please evaluate these training strategies separately with the architecture fixed to clarify their individual contributions.

  2. The Activity split contains overlapping training and test observations. The code first generates four-second windows with a 0.5-second stride and then randomly splits them, allowing the same raw observations to enter different sets. Reproducing this split shows that 99.16% of test target vectors appear in training histories and 47.03% in training labels. For example, all five targets in test window 8910 occur in training window 8912’s history. The resulting evaluation therefore does not establish performance on unseen future segments.

6.The training objective differs between the paper and the released code. Equation (7) includes only the forecasting loss and the auxiliary event loss, whereas the released code enables intermediate-view supervision LmidL_{\mathrm{mid}} and an EMA-teacher guidance loss LguideL_{\mathrm{guide}} by default from epoch 11 onward. Please clarify whether these losses were used for the main results.

7.The availability of future queries needs clarification. The prediction times, target variables, and number of outputs are determined by the supplied queries, which are constructed from actual future observations in the experiments. Please explain how these query locations would be obtained at deployment and how the method would be used when future observation times and counts are unknown.

Questions for authors

see Critical Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7LU8

Response to Reviewer 7LU8


Thank you for highlighting why local observations are worth keeping, why sparse histories make it hard to choose which information to trust, and how the dual-memory design fits these problems. We address W1–W7 below.


W1. What distinguishes SEQNet from existing aggregation, query decoding, and gated fusion?

What is new in SEQNet? As we explain in the Introduction and illustrate in Figure 1, compressing a history can hide local changes that matter for forecasting. As you noted, the dual-memory design keeps those details accessible. SEQNet retains selected observations alongside the compressed history so it can retrieve the details relevant to the variable and time being forecast.

Aggregation, query-based retrieval, and routing each serve that goal. The summary captures the overall trend. The retained observations preserve details such as when a rainfall peak occurred and how large it was. The query tells the model which forecast it is making, so it can look up the relevant details. When the target variable has few or outdated observations, the router adjusts how much the model relies on that variable’s history, other variables’ histories, and a learned prior (§§3.1, 3.3–3.4). Together, these components help SEQNet make the most of its core design by turning the local details it preserves into evidence for each forecast.

Does it matter which observations we keep? Figure 3(b) and Tables 9 and 10 compare event-retention strategies on USHCN.

USHCN configurationTest MSE
Full selection rule0.495000
Local change only0.500841
Deviation only0.505400
Recency only0.505680
Same number of randomly selected observations0.511604
Remove retained events0.501103

Removing the retained events raises MSE from 0.495000 to 0.501103. Replacing them with the same number of randomly selected observations raises MSE to 0.511604. At the same event budget, combining recency, deviation, and local change also works better than using any one of them alone. Which observations the model keeps makes a difference.

Does retrieving observations for each forecast help? Figure 3(a) and Table 7 add the components one at a time on PhysioNet. Each reduction is measured against the same reference configuration.

PhysioNet configurationMSE reduction
Soft Temporal Memory2.19%
Add retained events3.05%
Add query-based retrieval3.70%
Full SEQNet4.35%

Adding retained events increases the MSE reduction from 2.19% to 3.05%. Letting the model retrieve information based on the variable and time being forecast increases it further to 3.70%. Table 8 also shows that removing explicit observation-support information from the router raises MSE by 1.63% on USHCN and 0.32% on MIMIC-III.

In the main PhysioNet comparison (Table 2), SEQNet achieves an MSE of 0.00367, 24.49% lower than Hi-Patch (0.00486), the strongest baseline in that table. The ablations show why the core design helps. Keeping selected local observations reduces forecasting error, and retrieving information for the specific forecast reduces it further. We will link these two design choices directly to the corresponding ablations in the Introduction and §3.1.


W2. Recent baselines and evaluation protocol

We added APN, ASTGI, and HyperIMTS. Data construction, normalization, prediction targets, and metric reductions for PhysioNet, MIMIC-III, and USHCN follow §4.1 and Appendix E. Activity uses the subject-disjoint split described in W5.

MethodPhysioNetMIMIC-IIIActivity (subject-disjoint)USHCN
SEQNet0.00367 ± 0.000050.0920 ± 0.00060.0022320.495 ± 0.005
APN0.0040600.1014210.0026810.558340
ASTGI0.0037820.0987970.0023320.527501
HyperIMTS0.0049500.2033000.0025300.549600

SEQNet still achieves the lowest reported MSE on all four datasets after adding APN, ASTGI, and HyperIMTS. Compared with the strongest added baseline on each dataset, SEQNet reduces MSE by 4.29% on Activity with held-out subjects and 6.16% on USHCN. We will add these comparisons to §4.2 and document the evaluation protocol in Appendix E.


Due to the character limit, please see the remaining questions in the following comment.

Continued from previous comment


W3. Does the same event-selection strategy improve existing backbones?

We added the Mixed event selector to t-PatchGNN and Hi-Patch on USHCN. All three configurations for each backbone give equal weight to the losses from the original history and from a version with observations removed. They use the same five deletion families and four retention rates. The no-event configuration disables the event branch. Random and Mixed use the same backbone, training settings, number of retained events, and event branch. Only the choice of observations changes. Each configuration was run once (seed 1), and we selected the checkpoint with the lowest validation MSE on the original histories.

USHCN backbonePaired, no eventsPaired, random eventsPaired, Mixed events
t-PatchGNN0.6298720.6215520.611411
Hi-Patch0.6786030.6524600.625265

With the same number of retained events, Mixed lowers test MSE by 1.63% for t-PatchGNN and 4.17% for Hi-Patch compared with random selection. Compared with each backbone's paired run without events, MSE falls by 2.93% and 7.86%, respectively. We will add these results and explain how the event branch was added to each model in Appendix F. Table 2 reports the original benchmarks under the standard forecasting protocol.


W4. What does each training component contribute?

We kept the SEQNet architecture fixed on PhysioNet. Paired-view training predicts the same targets from the original history and from a version with observations removed (§3.5, Eq. (5)). Both weighting schemes sample deletion families uniformly. Constrained weighting gives harder families more weight in the forecasting loss, within the limits in Eq. (6). The auxiliary event loss uses binary cross-entropy to predict whether a target falls below the training-set 5th percentile or above the 95th percentile (§3.6, Eq. (7)).

ConfigurationPaired viewDeletion-family weightingEvent supervisionTest MSE
D0Off—Off0.003801
D1OnUniformOff0.003748
D2OnConstrainedOff0.003736
D3OnUniformOn0.003706
D4OnConstrainedOn0.003689
D5Off—On0.003787

Adding the paired view with uniform weighting reduces MSE by 1.39% without event supervision (D0→D1) and 2.14% with it (D5→D3). Within paired-view training, constrained weighting reduces MSE by 0.32% (D1→D2) and 0.46% (D3→D4). Event supervision reduces MSE in all three matched comparisons, by 0.37% (D0→D5), 1.12% (D1→D3), and 1.26% (D2→D4). D4 has the lowest MSE in this ablation, 2.95% below D0. We will add the results to Appendix F and identify the loss terms corresponding to each component in §§3.5–3.6.


W5. Overlapping windows in Activity

As your check shows, randomly splitting the original Activity windows allowed raw observations to appear in both training and test data. We now split the 25 subjects into training, validation, and test sets (15/5/5) before creating four-second windows at a 0.5-second stride. This yields 6,598, 2,001, and 2,140 windows, respectively. No subjects or raw observations are shared across splits. We fit normalization statistics using only the training subjects.

MethodSubject-disjoint test MSE
SEQNet0.002232
ASTGI0.002332
Hi-Patch0.002347
t-PatchGNN0.002511
HyperIMTS0.002530
APN0.002681

SEQNet retains its lead over all five baselines on held-out subjects with zero cross-partition observation overlap. It achieves 4.29% lower MSE than ASTGI, the strongest baseline in this comparison, and 4.90% and 11.11% lower MSE than Hi-Patch and t-PatchGNN, respectively. We will use the new results in §4.2 and Table 2, update the counts and split description in Table 1 and Appendix E.1, and release the subject lists and split code.


W6. Training objective in the paper and code

We tried intermediate-view supervision and EMA-teacher guidance early in development. They did not consistently improve validation performance, so we left both out of the final training setup. An old launcher that enabled them by default was mistakenly included when we packaged the code. The reported main results used Eq. (7) with both components disabled. We will align the released code with Eq. (7) and provide the corresponding training configurations and commands.

To make it easier to reproduce our results, we provide the code in an anonymous repository.


Due to the character limit, please see the remaining questions in the following comment.


Continued from previous comment


W7. How are future queries specified at deployment?

In our experiments, the test records tell us which variables to forecast and at what times. For example, on PhysioNet, if a heart-rate measurement was recorded at hour 30, we ask SEQNet to predict heart rate at hour 30 using the first 24 hours of history. The model receives the requested time and variable. The measured heart rate is used to score the prediction (§3 and Appendix E.1).

At deployment, the user or application specifies what to forecast and when. If the measurement schedule is known, it can supply the query times. If the schedule is unknown, the application can choose its own times. For example, it can request heart-rate forecasts 15, 30, 45, and 60 minutes ahead. SEQNet returns four forecasts from the history available at the time of the request, even if no measurement is eventually taken at those times.

An actual measurement is needed only to check how accurate a forecast was. If a requested time has no recorded value for that variable, the forecast can still be returned and is excluded from the test MSE. We will clarify in §3 and §4.1 how benchmark queries are chosen and how an application can specify its own queries.


Thank you very much for your constructive and valuable questions, which helped us improve the manuscript. We hope the additional experiments and clarifications address your concerns. If you have any other questions or suggestions, we look forward to continuing the conversation!


Comments on the paper

2 comments
Area_Chair_xHFsOpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

AuthorsOpenReview ↗

Submission 21774 — Review and Discussion Summary

First, we sincerely thank the AC for promoting the discussion, and the three reviewers for their careful evaluation during the review and rebuttal periods. After the discussion, hUB8 raised their score from 2 (weak reject) to 3 (weak accept). MKu8 maintained a score of 3, and 7LU8 maintained a score of 2.

ReviewerScoreStrengths highlightedMain concerns and our responseDiscussion outcome
7LU82Identified two worthwhile problems in irregular forecasting and found the dual-memory design aligned with them.Clarified how SEQNet combines global summaries, query-specific event retrieval, and routing based on the available observations. Added recent baselines and fixed-budget event comparisons. Clarified the training objective and future queries, and reran Activity with a recording-disjoint split.Maintained the score
hUB82 → 3Highlighted the design for sparse or stale histories, the four-dataset results, and the matched-budget event ablations.Added APN, ASTGI, and HyperIMTS, documented the evaluation protocols, compared strong baselines under the same deletion conditions, and isolated the training components.Said the response addressed most concerns and raised the score.
MKu83 → 3Highlighted the combination of global memory, local events, and support-aware routing, as well as the five-seed main results.Added observation-deletion comparisons with Hi-Patch and t-PatchGNN and ablated paired-view training, deletion-family weighting, and auxiliary event supervision with the architecture fixed.Maintained the score

Key evidence added during discussion

  • Recent baselines and the contribution of event selection. SEQNet achieved lower MSE than APN, ASTGI, and HyperIMTS on all four datasets. We will report these comparisons in §4.2. On USHCN, we also added the same event branch to t-PatchGNN and Hi-Patch. With the input adapter, training setup, and number of retained events fixed, selecting events with SEQNet’s rule reduced MSE by 1.63% and 4.17%, respectively, relative to random selection. This comparison tests whether which observations are retained matters when the event budget is fixed.
  • Robustness under matched training and test conditions. On PhysioNet, we trained t-PatchGNN and Hi-Patch with paired deletion and evaluated all three models on the same 20 deletion settings. SEQNet achieved mean and worst-case MSEs of 0.004503 and 0.006743, both lower than either paired-training baseline. In the fixed-architecture ablation, paired views provided the largest individual gain. Using all three training components reduced clean MSE by 2.95% relative to using none in that run.
  • Activity evaluation. The original protocol randomly split overlapping windows. In the rerun, we split the 25 recordings into 15/5/5 before generating windows. This prevents overlapping windows from the same recording from crossing partitions.
  • Training objective and reproducibility. We clarified that the main results used Eq. (7) with intermediate-view supervision and EMA-teacher guidance disabled. An older launcher enabled both by default. We will align the released launcher with the reported setting and provide the corresponding training commands.

Finally, thank you for taking the time to read our paper and consider our rebuttal!

All authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.