- 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
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.
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 task | AKP LC / LL | PatchTST LC / LL |
|---|---|---|
| HumanActivity, 3000 -> 600 ms | 0.04817 / 0.04647; 0.12387 / 0.12313 | 0.05185 / 0.04935; 0.13138 / 0.12789 |
| HumanActivity, 3000 -> 900 ms | 0.05240 / 0.05121; 0.13162 / 0.13140 | 0.05703 / 0.05438; 0.13893 / 0.13585 |
| HumanActivity, 3000 -> 1000 ms | 0.05413 / 0.05295; 0.13451 / 0.13402 | 0.05853 / 0.05609; 0.14086 / 0.13824 |
| USHCN, 24 -> 6 months | 0.54937 / 0.55065; 0.25892 / 0.25683 | 0.62093 / 0.61497; 0.36169 / 0.34774 |
| USHCN, 24 -> 12 months | 0.55737 / 0.55117; 0.26247 / 0.25746 | 0.65566 / 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.
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.
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-6patches. Complementary groups sum exactly to total target-cell SSE/SAE change0.28519/0.28521and0.31448/0.31487. The degenerate group contributes 99.9% of SSE worsening. The exactly-two-observation slice improves by0.00400MSE (2SE=0.00159)0.32120/0.33924and0.44255/0.44750. 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:
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
0.31369/0.30948and0.43090/0.42720; 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 proofRevision. 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