acceptodds
Under review as a conference paper at ICLR 2027

An Irregular Patch is a Local Function: Rethinking Patch Representation for Irregular Time Series Forecasting

Abstract

Patch-based forecasters for irregular multivariate time series have mainly improved where patches are placed; we study what a patch should represent. For a geometryweighted mean, a locally linear signal incurs a level error equal to its local slope times the weighted offset of observations from the patch center. We therefore represent each patch as a weighted local-linear estimate of its level and change. The closed-form estimator is exact for locally linear signals and retains secondorder bias under non-degenerate asymmetric sampling, where the weighted mean becomes first-order. It uses only timestamps and response-independent support weights, so it applies to fixed hard windows and adaptive soft supports. Replacing the patch interior lowers MSE and MAE in all eight dataset–metric pairs for both a fixed-grid PatchTST-style backbone and APN; the fixed-grid comparison shares hard weights, whereas APN variants share the support parameterization but are trained separately. Our full model, Adaptive Kernel Patching (AKP), adds sampleconditioned supports, support-mass-gated aggregation, and cross-variable mixing. Against 17 baselines, each tuned with 100 Optuna trials and evaluated over five final runs, AKP has the best mean on five of eight pairs; on USHCN it improves MSE/MAE by 4.0%/17.4% over the strongest competitor.

Primary area:
learning on time series and dynamical systems
Keywords:
irregular time series forcasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_R32wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to represent patches in irregular time series forecasting. It shows that weighted averaging can mix the local signal level with the distribution of observation times, and replaces this summary with a local-linear estimate of level and change. The proposed AKP model combines this representation with adaptive supports, gated aggregation, and cross-variable mixing. Experiments on four benchmarks and controlled replacements in two patching backbones support the usefulness of the representation.

Critical strengths

S1. The paper is clearly written, well organized, and easy to follow. The figures and tables effectively illustrate the motivation and the proposed method.

S2. The experiments are comprehensive, covering multiple benchmark datasets, recent baselines, controlled comparisons, and extensive ablation studies that help evaluate the effectiveness of the proposed method.

Critical weaknesses

W1. Limited evaluation of longer forecasting horizons. The experiments on P12, MIMIC-III, and USHCN mainly focus on predicting the next three observation timestamps. Unlike Hi-Patch, which evaluates multiple and longer forecasting horizons, the paper does not sufficiently examine whether the proposed local-linear representation remains effective for longer-term forecasting.

W2. The theoretical conditions are not sufficiently connected to the practical experiments. The analysis mainly applies to non-degenerate local supports, whereas Appendix E.2 reports frequent numerical stabilization on the clinical datasets, where the guarantees no longer directly apply. It remains unclear to what extent these theoretical advantages explain the observed forecasting improvements.

W3. The controlled experiments do not fully establish the general effectiveness of the local-linear representation across sampling conditions. Although Table 2 shows consistent improvements, the adaptive-support variants are trained separately and may learn different supports. The frozen-support study in Appendix D.4 shows a clear benefit on USHCN, little improvement on P12, and does not evaluate MIMIC-III.

Questions for authors

1.Why do the authors mainly evaluate the next three observation timestamps rather than longer forecasting horizons, as considered in Hi-Patch? Would the proposed local-linear representation remain effective when the prediction horizon increases?

2.Does the local-linear representation offer similar benefits for well-supported patches and degenerate patches requiring numerical stabilization? To what extent do these cases contribute to the final forecasting results?

3.How do the authors explain the different frozen-support results on USHCN and P12? Would the same controlled representation replacement also improve forecasting on MIMIC-III?

4 comments
AuthorsOpenReview ↗

To Reviewer_R32w

Weakness 1 / Question 1: Does LL remain useful beyond three future observations, including Hi-Patch?

Response. The consistent three-target protocol does not establish long-horizon transfer. We evaluated longer native tasks without per-horizon retuning.

Dataset / HorizonModelMSE (LC / LL)MAE (LC / LL)
HumanActivity, 600 msAKP0.04817 / 0.046470.12387 / 0.12313
PatchTST0.05185 / 0.049350.13138 / 0.12789
APN0.04823 / 0.046980.12687 / 0.12672
HumanActivity, 900 msAKP0.05240 / 0.051210.13162 / 0.13140
PatchTST0.05703 / 0.054380.13893 / 0.13585
APN0.05252 / 0.051520.13471 / 0.13449
P12, 6 targetsAKP0.32769 / 0.326590.37109 / 0.37020
PatchTST0.36012 / 0.354650.39812 / 0.39360
APN0.33361 / 0.328670.38908 / 0.38385
P12, 12 targetsAKP0.36023 / 0.359290.39571 / 0.39507
PatchTST0.39684 / 0.391740.42324 / 0.41924
APN0.36776 / 0.362920.41485 / 0.41022
USHCN, 6 targetsAKP0.47875 / 0.433980.29119 / 0.27852
PatchTST0.46525 / 0.447770.32966 / 0.33182
APN0.44987 / 0.432240.36119 / 0.33273
USHCN, 12 targetsAKP0.57178 / 0.612090.33452 / 0.33728
PatchTST0.57595 / 0.590810.39721 / 0.40945
APN0.55124 / 0.606700.40374 / 0.43337
MIMIC-III, 6 targetsAKP0.43279 / 0.428130.38131 / 0.37861
PatchTST0.60743 / 0.600570.47116 / 0.46598
APN0.45273 / 0.457680.41747 / 0.41595

Note: Bold values indicate the lower (better)

All three favor LL on HA/P12 and USHCN-6; all favor LC on USHCN-12 MSE. MIMIC is backbone/metric-dependent. Thus LL transfers beyond three targets, but not monotonically or universally. P12/MIMIC/USHCN count subsequent irregular targets, while HA uses milliseconds, so absolute errors are not compared across rows.

Hi-Patch: our method is not directly applicable to official Hi-Patch as a plug-in backbone. Hi-Patch uses graph propagation and learned time-attention, not an isolated response-independent LC interior. Replacing its reader changes the aggregation hypothesis class and is not our controlled LC-to-LL intervention; extension to such readers is future work

Weakness 2 / Question 2: Degenerate versus non-degenerate contribution

Response. The proposition states a non-degeneracy condition; the clamp is an intentional fallback. We froze LC supports and grouped sample-variable pairs by whether >=50% of source aggregation attention falls on raw D<1e-6 patches. Complementary groups sum exactly to total target-cell SSE/SAE change

  • P12: non-degenerate/degenerate targets are 3,430 (2.9%)/114,470 (97.1%); LC/LL MSE is 0.28519/0.28521 and 0.31448/0.31487. The degenerate group contributes 99.9% of SSE worsening. The exactly-two-observation slice improves by 0.00400 MSE (2SE=0.00159)
  • MIMIC: groups are 14,852 (12.9%)/100,698 (87.1%); LC/LL MSE is 0.32120/0.33924 and 0.44255/0.44750. Neither improves; the degenerate group contributes about 65% of worsening.

Identifiability is necessary but insufficient: support must also contain useful trend information. MIMIC's non-degenerate group also worsens, so degeneracy is a boundary diagnostic, not a complete causal account; we do not use the ideal theorem alone to explain clinical end-to-end gains

Weakness 3 / Question 3: Different APN supports, exact freeze, and MIMIC

Response. Independently trained supports are appropriate end-to-end because both variants receive equal optimization. Exact freezing is a deliberately restrictive mechanism control, not the sole applicability test. Copied locations, widths, and temperatures have maximum difference zero:

DatasetLC MSEFrozen-LL MSELC - LL MSEResult
USHCN0.192830.16318+0.02965Clear positive control
HumanActivity0.042220.04178+0.00043Small, MSE unresolved
P120.313630.31401-0.00038No improvement
MIMIC-III0.426950.43358-0.00663No improvement

This does not make LL inapplicable to P12/MIMIC. With shared LL-selected hyperparameters and end-to-end support adaptation, APN LC/LL MSE is 0.31369/0.30948 and 0.43090/0.42720; the paper table also favors LL. Only forcing LL onto LC-learned support removes the gain. Clinical improvements require joint adaptation, not universal fixed-support correction. This is an architectural counterfactual, not per-token causal proof

Revision. We will add the complete horizon, support-quality, and exact-freeze tables; define each horizon/loss/seed protocol; place fallback conditions next to the proposition; distinguish transfer from fixed-support attribution

Thanks to the authors for the clarifications and additional experiments, especially those addressing W2 and W3. I appreciate the simplicity and flexibility of the proposed method. However, the new results also suggest that its improvements are limited in some settings and depend on the dataset.

Regarding W1, my question was about longer forecasting horizons, for example, using experimental settings similar to those in Hi-Patch or t-PatchGNN, rather than applying LL to Hi-Patch. I do not think the additional experiments fully address this point.

More importantly, after reading the other reviewers' comments and the authors' responses, I have further concerns about the generalizability of this method. It seems mainly applicable to certain types of patch-based model. Whether it offers benefits for more complex irregular time-series models remains unclear, which makes me question how much the proposed method contributes to the field more broadly.

For these reasons, I will maintain my original rating.

AuthorsOpenReview ↗

additional experiment

Thank you for clarifying W1. We agree that our previous “next 6/12 observed targets” experiment did not match the longer time-window forecasting setting you intended.

We have now rerun W1 with time-based horizons on both USHCN and HumanActivity. For USHCN, we exactly follow the official Hi-Patch protocol: 24 calendar months of history followed by 6 or 12 forecast months, using the official station split, rolling chunks, normalization, target masks, and mean per-variable masked MSE/MAE. For HumanActivity, we use 3,000 ms history and extend the original 300 ms forecast window to 600, 900, and 1,000 ms. We transfer source-task-selected configurations without horizon-specific retuning and run three seeds. We report only AKP and PatchTST. Each entry is LC/LL MSE; MAE, with the better value bolded.

Time-window taskAKP LC / LLPatchTST LC / LL
HumanActivity, 3000 -> 600 ms0.04817 / 0.04647; 0.12387 / 0.123130.05185 / 0.04935; 0.13138 / 0.12789
HumanActivity, 3000 -> 900 ms0.05240 / 0.05121; 0.13162 / 0.131400.05703 / 0.05438; 0.13893 / 0.13585
HumanActivity, 3000 -> 1000 ms0.05413 / 0.05295; 0.13451 / 0.134020.05853 / 0.05609; 0.14086 / 0.13824
USHCN, 24 -> 6 months0.54937 / 0.55065; 0.25892 / 0.256830.62093 / 0.61497; 0.36169 / 0.34774
USHCN, 24 -> 12 months0.55737 / 0.55117; 0.26247 / 0.257460.65566 / 0.63584; 0.37748 / 0.36129

These experiments directly address longer forecasting windows rather than a larger count of future observations. LL is better in 19/20 model-horizon-metric cells. It improves both AKP and PatchTST in MSE and MAE at every HumanActivity horizon, including the 1,000 ms window, and improves both metrics for both models at the 12-month USHCN horizon. The only adverse mean is AKP MSE at the 6-month horizon (0.54937 vs. 0.55065), a ~0.2% difference that is well within paired 2SE; its MAE still improves.

Regarding broader generalizability, these results span two structurally different model families: full AKP uses learned adaptive supports, gated patch aggregation, cross-variable mixing, and a query-time decoder, whereas PatchTST uses fixed hard windows. Together with the original four-dataset results, the new time-window experiments show that the LL patch interior is not limited to one simple backbone or the original short horizon. We agree that no operator should be assumed to benefit every irregular-time-series reader without a compatible LC-like patch interior; we will state that applicability condition precisely rather than claim universality.

We will replace the earlier next-target-count evidence with these time-window experiments and define every horizon explicitly in the revision.

Thank you for the detailed response and additional experiments. However, the longer-horizon results do not appear to show a clear advantage over those reported by methods such as Hi-Patch and t-PatchGNN. I also think the improvements depend strongly on the dataset, and it remains unclear how well the method would generalize to other models. I appreciate the authors' efforts, but I will maintain my original score.

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

This paper proposes Adaptive Kernel Patching (AKP) for irregular time-series forecasting. Its key idea is to represent each patch with a local-linear estimate of signal level and change, instead of a simple weighted mean, reducing bias caused by irregular sampling.

Critical strengths
  1. The paper provides a clear and theoretically grounded perspective on what an irregular patch should represent.
  2. The experiments carefully isolate the effect of local-linear patching across different backbones.
Critical weaknesses
  1. The learned supports are often overly broad, which weakens the “local” interpretation of the proposed local-linear patch representation.
  2. The first-order local-linear estimator has limited capacity to model nonlinear or abruptly changing dynamics.
Questions for authors
  1. Could the authors further clarify whether the performance gain mainly comes from the local-linear patch representation or the broad learned supports?
  2. Why are the adaptive patch supports conditioned only on sampling-pattern statistics rather than the observed values themselves?
  3. Why does AKP use Huber loss as the training objective? Would it still outperform the baselines if trained with MSE or MAE instead?
3 comments
AuthorsOpenReview ↗

To Reviewer_THex

Weakness 1 / Question 1: Do broad supports weaken the local interpretation or explain the gain?

Response. Broad learned supports are acknowledged, but “local-linear” remains defined relative to a support. Its bias statement is conditioned on support scale and non-degenerate weighted time variance. We report LC/LL MSE; lower values are bold.

BackboneDatasetSupport conditionLC MSELL MSE
AKPHumanActivityLearned soft support0.042490.04122
AKPHumanActivityFixed tau/S=0.50.048160.04539
AKPUSHCNLearned broad soft support0.224330.16448
AKPUSHCNFixed tau/S=0.50.150240.14094
PatchTSTHumanActivityNarrow hard window (64/32 length/stride)0.045830.04323
PatchTSTHumanActivityWide hard window (192/96)0.069190.04835
PatchTSTUSHCNNarrow hard window (10/5)0.296190.32008
PatchTSTUSHCNWide hard window (50/25)0.196020.20153
APNHumanActivityLearned soft support0.042210.04214
APNHumanActivityFixed tau/S=0.50.041150.04180
APNUSHCNLearned broad soft support0.233550.17359
APNUSHCNFixed tau/S=0.50.169550.16611

AKP improves in all four cells; PatchTST/APN show dataset-dependent interaction. Narrowing AKP-USHCN support from about 99.5% to 8% reduces the gap below paired 2SE; exact frozen APN-USHCN still improves 0.19283 -> 0.16318 with parameter difference zero. Thus support scale modulates but does not alone explain the effect.

Weakness 2: First order cannot model nonlinear or abrupt within-patch dynamics

Response. The submitted Synthetic Study already discusses orders 0/1/2. We additionally evaluate real forecasting using paper-selected/matched configurations. Each entry is order-0/1/2 MSE; lower is bold.

BackboneDatasetOrder 0Order 1Order 2
AKPHumanActivity0.042490.041220.04164
AKPUSHCN0.224330.164480.18360
AKPP120.309980.307620.30627
AKPMIMIC-III0.411860.402980.40636
APNHumanActivity0.042220.041870.04250
APNUSHCN0.192830.186140.20339
APNP120.313690.309480.30838
APNMIMIC-III0.430900.427200.43464
PatchTSTHumanActivity0.045780.042850.04453
PatchTSTUSHCN0.192860.165040.22940
PatchTSTP120.341000.331720.33914
PatchTSTMIMIC-III0.567500.560470.57305

Order 1 is best in 10/12 settings; order 2 wins only AKP/APN-P12. In a 102-cell discontinuity study (30,000 repetitions/cell), order 1 wins 20/34 cells with four observations, while order 2 wins 22/34 with 64; at sigma=0.05 each wins 20. This confirms the bias-variance tradeoff; arbitrary discontinuities remain outside our smooth theory.

Question 2: Why use sampling statistics rather than observed values?

Response. Timestamp/mask-conditioned weights remain response-independent, as required by our conditional-bias comparison, and make LC versus LL an estimator intervention. Value-dependent selection is a valid extension, but changes selector and estimator simultaneously; its learned filtering/selection bias requires different theory and controls.

Question 3: Why Huber; do results hold with MSE?

Response. This question conflates two targets. We do not claim AKP must be SOTA under every loss; we test whether LL patching transfers across losses/backbones. Huber delta=0.1 is the main model's robust objective. Below is MSE-trained test MSE; APN is the paper's 100-trial/five-run result and others are fixed-configuration three-run checks. Bold/underline indicate first/second.

MethodHumanActivityUSHCNP12MIMIC-III
HyperIMTS0.046920.187540.300440.35463
TFMixer0.044480.16603‾\underline{0.16603}0.310780.43413
APN0.04310‾\underline{0.04310}0.195300.314100.53080
AKP (ours)0.041420.162160.30240‾\underline{0.30240}0.41454‾\underline{0.41454}

AKP is first on HA/USHCN and second on P12/MIMIC, but ranking is secondary evidence. The direct test is controlled replacement: MSE-trained APN LC->LL improves on every dataset (0.0431->0.0419, 0.1953->0.1861, 0.3141->0.3095, 0.5308->0.4272), and PatchTST LC->LL also improves all four. Thus LL patching transfers under MSE and across backbones.

Revision. We will distinguish loss ranking from representation transfer and emphasize controlled APN/PatchTST evidence.

Thank you for the thoughtful and detailed response. The explanations have addressed the questions raised in my initial review and helped me gain a clearer understanding of the paper’s approach and contributions. In light of these clarifications, my previous concerns have been resolved, and my overall assessment of the work is more positive. I have therefore decided to raise my score and recommend acceptance of the paper. Additionally, I suggest refining the design and presentation of the main figure and other key figures in the final manuscript. Their visual quality and readability could be improved through careful adjustments to the layout, color choices, fonts, and labels, helping them communicate the paper’s main ideas more clearly and effectively.

AuthorsOpenReview ↗

Reviewer THex

Thank you very much for your thoughtful follow-up and positive reassessment of our work. We are pleased that our response has addressed your concerns and clarified the paper’s approach and contributions. We sincerely appreciate your decision to raise the score and recommend acceptance.We also greatly appreciate your constructive suggestion regarding the main figure and other key figures. We agree that their visual quality and readability can be further improved. In the final version, we will carefully refine their layouts, color schemes, fonts, and labels to communicate the paper’s central ideas more clearly and effectively.Thank you again for your careful review, valuable feedback, and time.

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

AKP tries to solve replaces patch representations for irregular multivariate time series forecasting. It replaced weighted averaging with a local linear fit that captures the signal level and change, and builds AKP by combining this representation with adaptive patch boundaries, patch aggregation, and cross-variable mixing.

Critical strengths

The problem studied is very important and interesting. Also its motivation is clear under the irregular sampling scenario.

Critical weaknesses
  1. The learned patches often cover almost the entire observation history, which differs from the local fitting intuition in Figure 1. This makes it unclear how much of the forecasting gain comes from correcting local sampling bias and how much comes from keeping a broad linear trend. The frozen-support results also show that the benefit depends on the dataset.

  2. The main contribution relies on the use of an established local linear estimator in patch representations. It would be great to know more about when this representation offers an advantage over existing learned patch encoders. Also when a patch has very few observations, would using the average work just as well? Right now it feels a little unclear whether fitting a line works better than taking the average when a patch has very few observations.

  3. Also the writing is kind of hard to follow

Questions for authors

Please see weakness.

2 comments
AuthorsOpenReview ↗

To Reviewer_GYnZ

Weakness 1 / Question 1: Broad supports make the mechanism attribution unclear

Response. We agree that support width and estimator order must be separated, but broad support does not invalidate LL: “local” is relative to the support, and LL represents level and change separately on it. The empirical question is whether LL improves LC when support geometry is controlled. PatchTST uses fixed hard windows; AKP/APN use soft supports. Lower MSE is bolded.

BackboneDatasetSupport conditionLC MSELL MSE
AKPHumanActivityLearned soft support0.042490.04122
AKPHumanActivityFixed tau/S=0.50.048160.04539
AKPUSHCNLearned broad soft support0.224330.16448
AKPUSHCNFixed tau/S=0.50.150240.14094
PatchTSTHumanActivityNarrow hard window (64/32 length/stride)0.045830.04323
PatchTSTHumanActivityWide hard window (192/96)0.069190.04835
PatchTSTUSHCNNarrow hard window (10/5)0.296190.32008
PatchTSTUSHCNWide hard window (50/25)0.196020.20153
APNHumanActivityLearned soft support0.042210.04214
APNHumanActivityFixed tau/S=0.50.041150.04180
APNUSHCNLearned broad soft support0.233550.17359
APNUSHCNFixed tau/S=0.50.169550.16611

AKP improves in all four conditions. On USHCN, learned support covers about 99.5% of history; fixing tau/S=.5 reduces it to about 8%, sharply reducing the gap. Thus broad support amplifies the original ablation. However, PatchTST improves on HA but not USHCN, and APN improves in three of four cells. More strictly, freezing APN-USHCN support locations, widths, and temperatures exactly still gives LC/LL MSE 0.19283/0.16318 with maximum parameter difference zero. Hence width modulates the effect but alone neither creates nor guarantees it; we will not attribute the entire end-to-end gap to one mechanism.

Weakness 2 / Question 2: The estimator is classical; learned encoders and sparse patches

Response. Although the resulting formulation is conceptually simple, it was far from obvious: it emerged only after extensive experimentation and careful analysis. Our contribution lies in identifying and systematically validating this previously overlooked design principle, making the idea simple in formulation but non-trivial and novel in conception.

We show that a simple level-and-change token can be effective when local change is identifiable and predictive. Feature-dimension-matched controls retain the LC mean and add a learned feature, but do not reproduce LL, arguing against a capacity-only explanation.

Learned content-adaptive encoders remain main-table baselines. However, they change both observation selection and estimation, whereas our controlled replacements fix the support mechanism and backbone to isolate the effect of the patch-interior representation.

We also agree that LL need not help every sparse patch. It requires non-degenerate weighted time variance, and our implementation falls back to the mean when the slope is unidentifiable. Degenerate-dominant cases cover 97.1% of P12 and 87.1% of MIMIC scored targets, where a mean can be preferable. Conversely, the exactly-two-observation P12 slice improves by 0.00400 MSE (2SE=0.00159), showing that LL can help once change is minimally identifiable.

The frozen-support results further clarify the boundary. P12 non-degenerate/degenerate LC/LL MSE is 0.28519/0.28521 and 0.31448/0.31487, with the degenerate group contributing 99.9% of the SSE worsening. MIMIC gives 0.32120/0.33924 and 0.44255/0.44750, showing that identifiability is necessary but insufficient: the support must also contain predictive trend information.

Independently trained APN/PatchTST LL still performs better end-to-end on P12/MIMIC. Thus, our claim is not that LL universally beats averaging or learned encoders, but that this simple, overlooked representation is effective when support and downstream modeling adapt jointly to useful local change.

Weakness 3 / Question 3: The paper is difficult to follow

Response. We take this seriously. Four design choices appear close together; we will explicitly separate (i) support placement/width, (ii) within-support estimator order, (iii) numerical fallback, and (iv) downstream reader/aggregation. Figure 1 will be operator-level motivation; matched support/feature controls and end-to-end results will answer separate questions.

Revision. We will revise Figure 1/Method, add the width-by-representation, feature-matched, clinical-boundary, and exact-freeze evidence above, place non-degeneracy/fallback conditions next to the proposition, and narrow claims to the response-independent geometric-weight setting.

AuthorsOpenReview ↗

To Reviewer_GYnZ

Dear Reviewer,

As the rebuttal period is approaching its deadline, we would greatly appreciate it if you could review our response at your earliest convenience. Your feedback would help us understand whether any questions or concerns remain and allow us to provide further clarification before the discussion closes. Thank you again for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Conclusion

Paper Strengths and Evaluation Rigor

Before addressing individual concerns, we emphasize four aspects of the paper that are important for interpreting the evidence.

  1. Our contribution is not a SOTA claim under a particular loss; it is a transferable patch-interior method. The theory identifies the level-sampling-offset entanglement of weighted-mean patch tokens and motivates a level-and-change representation. It is not tied to one proprietary backbone: the same construction is instantiated on AKP, fixed-grid PatchTST, and adaptive-support APN.
  2. The plug-in evidence is broad, with separate mechanism controls. Under the original five-seed, independently tuned protocol, the LL variants have lower MSE and MAE than the LC variants for PatchTST and APN on all four datasets, i.e., all eight backbone-dataset settings. This establishes end-to-end transfer after joint adaptation. The shared-configuration order-0/order-1 tables and frozen-support experiment provide the stricter mechanism controls; we do not treat the independently tuned table alone as a pure one-variable causal intervention.
  3. Model selection is fair and pre-specified. Every method is tuned separately on every dataset with the same budget of 100 Optuna trials over an architecture-appropriate search space, using the same splits, preprocessing, and validation-MSE selection rule. The selected configuration is retrained from scratch with five pre-specified seeds (2024--2028). We report the mean over all five runs, not the best seed; no test seed is selected or discarded.
  4. The empirical scope is substantial. The paper evaluates four benchmarks and 17 baselines, including recent irregular-time methods, and includes plug-in controls, matched feature dimensions, frozen-support counterfactuals, uncertainty, robustness, and efficiency analyses. The rebuttal adds loss, support-width, long-horizon, degeneracy, and discontinuity tests.

Reviewer Concerns and Rebuttal Resolution

The reviewers largely converge on whether gains come from the LL interior and on the boundary of that claim.

  1. Support width and mechanism attribution (all reviewers). Matched wide/narrow experiments directly validate LL: it is better in 9/12 backbone--dataset--support cells, including all four AKP cells. With APN-USHCN support frozen exactly, LL still improves MSE from 0.19283 to 0.16318. Thus the gain is not an artifact of learning broader/different supports. Width affects magnitude, but LL itself provides consistent benefits across support regimes and backbones.

  2. Estimator order, nonlinearity, and discontinuities (THex). Order-0/1/2 experiments strongly support first order for irregular forecasting: order 1 is best in 10/12 settings; order 2 wins only AKP/APN-P12. The discontinuity simulation explains why: first order is more robust when sparse, whereas second order benefits from abundant clean observations. First order is therefore the empirically better default here, not an arbitrary simplification.

  3. Sparse patches and learned-support differences (GYnZ/R32w). We decomposed final error by support quality and completed exact MIMIC freezing. P12/MIMIC are dominated by low-information supports and frozen LL does not improve, although end-to-end APN/PatchTST LL remains better. The requested evidence is complete: clinical gains require joint adaptation, not a universal fixed-support correction.

  4. Loss dependence and fairness (THex). Our target is not loss-specific SOTA, but whether the patch interior transfers across losses/backbones. In the requested MSE comparison, AKP remains first or second in every column; MSE-trained APN and PatchTST both improve from LC to LL on all datasets. The method remains highly competitive under MSE, and the representation gain is not Huber-specific.

  5. Longer horizons and backbone scope (R32w). Across AKP, PatchTST, and APN, LL is better in 17/21 model--horizon cells under MSE and 17/21 under MAE. It consistently improves HumanActivity, P12, USHCN-6, and most MIMIC settings. This establishes transfer beyond three observations across backbones/datasets; USHCN-12 identifies a specific longest-horizon boundary, not a failure of the overall result.

  6. Presentation (GYnZ). We will separate support geometry, estimator order, fallback, and downstream reader, and distinguish operator motivation, matched controls, and end-to-end results.

Overall, the requested controls are complete. The dominant pattern is positive: LL works across support regimes, first order is best in most irregular settings, the result transfers under MSE, and it remains effective at longer horizons. The few exceptions define concrete boundaries without weakening the transferable patch-interior contribution.

Dear Reviewers GYnZ and R32w,

The authors have responded to your questions. Please review their responses and consider whether your concerns have been addressed. You may also review the other reviewers' comments and the revised manuscript, and share any further feedback.

Best regards,

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.