acceptodds
Under review as a conference paper at ICLR 2027

Task-Aware Temporal Coarsening under Finite Budgets: Exact Risk, Tractable Surrogates, and Downstream Transfer

Abstract

Compressing an -step history into contiguous measurements requires deciding where to allocate temporal resolution. Standard heuristics allocate it uniformly or logarithmically, ignoring both the history covariance and the downstream task geometry , while the exact task-weighted compression risk is not block-additive and cannot be minimized exactly by the classical contiguous-partition dynamic program. We study task-aware temporal coarsening with and estimated on development data only: a dynamic program globally minimizes a block-additive surrogate , and its minimizer is refined under exact . This separates three gaps that are usually read as a single gap: approximation, optimization, and transfer. Over audit partitions per dataset–horizon cell at , surrogate rank fidelity ranges from on Weather to on Electricity. Exact orders six controlled constructions against a held-out linear reader at mean rank agreement , against for the surrogate, yet in a post-hoc validation audit, lower- partitions reduce neural-reader error in only six of thirteen cases. Under the linear reader, task-aware allocation beats both uniform and logarithmic spacing in all cells; under a Transformer reader, it beats uniform in all twelve cells, negligibly in one, but does not dominate a fixed logarithmic grid. Evaluating coarsened representations, therefore, requires auditing approximation, search, and transfer separately rather than attributing downstream performance to a single objective.

Primary area:
learning on time series and dynamical systems
Keywords:
Temporal compression, temporal coarsening, task-aware representations, task-weighted risk, dynamic programming, surrogate optimization, finite-budget optimization, representation transfer, time-series modeling.

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_5YHwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to compress a long time series into a small number of time intervals while keeping information useful for forecasting. The proposed method chooses the intervals using historical data and the prediction task. It also examines whether better compression scores lead to better predictions. Experiments on three datasets show improvements over equal-length intervals, but the method does not consistently beat a simple logarithmic spacing rule with a neural predictor.

Critical strengths
  1. The paper shows that a better compression score does not always lead to better predictions. This is useful because it explains why a method can look good mathematically but fail in practice.
  2. The experiments separate the choice of time intervals from the prediction model. This makes it easier to judge whether the intervals help, and the authors clearly report when their method loses to a simple baseline.
Critical weaknesses
  1. The method does not consistently beat simple logarithmic spacing with a neural predictor. This makes its practical advantage unclear, since the proposed method needs extra estimation and optimization.

  2. The experiments cover only three datasets and one neural predictor. It is unclear whether the findings hold for other prediction models.

Questions for authors
  1. When should users choose your method over simple logarithmic spacing, given its extra computation and mixed prediction results?

  2. Have you tested other neural predictors? If so, do they show the same gap between compression scores and prediction accuracy?

Reviewer_v58cOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates task-aware temporal coarsening, aiming to compress long time series into a fixed number of contiguous intervals while preserving information relevant to downstream forecasting. It formulates an exact task-weighted compression risk based on historical covariance and task geometry, and proposes a two-stage optimization procedure combining dynamic programming on a tractable surrogate with local refinement under the exact risk. Through theoretical analysis and experiments with linear and Transformer predictors, the paper examines approximation, optimization, and transfer gaps, revealing that lower analytical compression risk does not consistently translate into better neural forecasting performance.

Critical strengths
  1. Task-Aware Temporal Coarsening under Finite Budgets is an interseting idea.
  2. The experiment verified the effectiveness of the proposed method.
Critical weaknesses
  1. This paper has obvious problems in terms of clarity and presentationn. The introduction section fails to clearly explain the actual prediction task, the motivation for continuous-time aggregation, and the relationship between the proposed linear compression risk and the downstream neural network prediction performance. Instead, before delving into the mathematical details, theoretical conditions, optimization methods, and numerical results, the article introduces too much content, making it difficult for readers to form an intuitive understanding of the research problem and core challenges.
  2. The paper does not clearly distinguish whether its main contribution is an improved time coarsening algorithm or an analysis of the limitations of task-aware risk optimization.
  3. The proposed method does not consistently outperform a simple logarithmic allocation under the neural reader
  4. The experiments cover only three datasets, one neural architecture, and one neural compression budget.
  5. It would be best to provide some intuitive method diagrams to help readers understand the specific methodology and the core motivation.
  6. I believe that the presentation format of the paper should be more logical, so as to facilitate readers' understanding of the form of the task, the core motivation and challenges, as well as the unique ideas of this paper. But at present, it seems more like a disorganized course report.
  7. The author mainly compared some classic approaches. However, there was a lack of comparison with some mainstream deep learning models. Currently, many studies have been conducted on datasets such as ETT.
Questions for authors

See weakness

1 comment

Just as other reviewer stated, I am also concerned about the content of the paper. It doesn't seem to be at the level expected for an ICLR paper. I am inclined to reject it.

Reviewer_4n9qOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2(+1)
Presentation
2(+1)
Soundness
2
Read review
Paper summary

The paper places K ≪ H contiguous averaging bands over an H-step history, choosing boundaries from a task-weighted risk J(P; Σ, M) = tr(M R(P; Σ)), with Σ the history covariance and M = CᵀC the second moment of a ridge linear pilot, both estimated on TRAIN/VAL only, so the partition is frozen before any reader is trained. Because J is not block-additive under correlated Σ, boundaries come from a Fisher/Bellman DP on a block-diagonal-truncated surrogate J̃, then ±1 refinement on exact J. The audit covers ETTm1/Weather/Electricity (H = 512) with six arms under a linear reader (K ∈ {8, 16, 32}), three under a Transformer at K = 16 with three seeds, and a 20,000-partition J-vs-J̃ comparison. Task-aware allocation beats uniform under both readers but loses to a fixed log grid on two of three datasets under the Transformer, and the surrogate's fidelity collapses on Electricity; the paper reads this as three separable gaps — approximation, optimization, transfer.

Critical strengths
  1. The protocol is unusually guarded, which is what makes the negative findings usable. The ETTm1 pass/fail criterion was fixed before TEST and is reported as failed; fifteen pre-outcome gates ran before any result was read, three checking proved statements, two poisoning TEST to catch leakage; all 324 cells are reported, including the 0.046% “win” the authors call negligible. For a paper whose asset is a negative result, this is the right protocol.
  2. The J-vs-J̃-vs-reader instrument is reusable. Same geometry, same candidates, one variable: cross-block truncation. That J̃ is negatively rank-correlated with the held-out linear reader in 13 of 36 cells while J never is (Tables 9–10) warns anyone optimizing a decomposable surrogate.
Critical weaknesses
  1. The positive claim rests on the two comparators that are easiest to beat. uniform is weak by the paper's own numbers: an untouched log grid gains 16.4% / 8.0% / 24.4% over it. The informative comparison is log, where the Transformer gives +2.7% (ETTm1), +1.2% (Weather), −9.1% (Electricity); the ETTm1 losses at S = 192 and 336 (0.4726 vs. 0.4450; 0.5017 vs. 0.4640) are several times the seed spread. The 36/36 linear sweep is the least independent regime: J and the linear head regress on the same band means, and §5.2 concedes the set “is not fully neutral”. Also, log is one untuned geometric sweep, and “input only” equals “uniform” on ETTm1/Weather, with Weather at K = 16 ordering all six arms identically — so the 0.980 vs. 0.362 separation comes entirely from ETTm1 and Electricity. The paper's one new premise, that M carries information Σ does not, is therefore never tested under the nonlinear reader: task only / cov only / input only never reach it.
  2. The surrogate is never shown necessary, and the optimization gap has an uncalibrated reference. By Proposition 1, J̃(P) = J(P; B_P(Σ), M) is not an approximation of a fixed functional but the exact risk under a model that changes with P, and Proposition 2 shows the gap is sign-indefinite. The DP stage then contributes little: refinement removes a median 89.7% of the proposal's excess risk, the best of 32 random restarts beats the DP-initialized partition in 19 of 36 cells, and no compute parity is reported. The headline numbers — 27.5% excess, 89.7% recovery, “3.0% above best-found” — are all relative to a 23,041-candidate pool that is itself a search output, and only the DP, in gate G2, is checked against exhaustive enumeration. So “search limitations persist” cannot be separated from “the pool is search-limited”. At the H ≤ 18 draws of G2, K = 8 admits C(17,7) ≈ 19k enumerations, so the true ratio is computable.
Questions for authors
  1. Can you add a VAL-tuned geometric family (one decay ratio)? If a task-agnostic multiscale grid tuned on VAL captures most of the gain over uniform — especially the 9.1% on Electricity — what evidence remains that M̂, rather than a better-placed prior, is doing the work?
  2. How should this be positioned against the learned-boundary line (Pathformer, TimeSqueeze, TimeMosaic, Liu et al. 2026, Zucchi et al. 2026)? Your advantage — allocation fixed before the reader, hence transferable — is asserted, since §6 excludes end-to-end patching. If a boundary predictor were trained jointly with your own 2-block reader at a matched token budget, would “lower exact J does not imply lower reader error” and “log wins on ETTm1 and Weather” survive, or are they artefacts of fixing the partition early?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.