acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Patch Aggregation for Irregular Time Series Forecasting with Voronoi Quadrature

Abstract

Patching has become the efficient backbone of irregular multivariate time series (IMTS) forecasting, yet the step that turns scattered records into a patch token has remained largely unexamined. We revisit that step by formalizing a patch operator as a pair: a support kernel delimiting its temporal scope, and an aggregation measure assigning integration mass to the observations within that support. Prior work has increasingly refined the support kernel while leaving the aggregation measure at one unit of base mass per observation, so patch content estimates an average under the observation-induced measure rather than over physical time. We formalize this discrepancy and prove that it equals the patch-local covariance between temporal content and observation intensity, divided by the mean intensity. Being deterministic, the discrepancy resists averaging: it vanishes on a regular grid, grows whenever recording density correlates with content, and is fixed before any learnable module reads the token, beyond the reach of added capacity. Guided by this identity, we propose VQPatch, which treats every record as the representative of its Voronoi temporal cell, converts cell widths into quadrature weights, and retains the observation pattern as a bounded prediction condition. The resulting token provably approximates the physical-time patch integral at a rate governed by the cell mesh and independent of the sampling intensity, at no asymptotic cost beyond sorting. Across four IMTS benchmarks and seventeen baselines, VQPatch attains the lowest MAE on every dataset, improves MSE over the strongest patch baseline by up to 9.36%, and cuts peak training memory by 61.4%.

Primary area:
learning on time series and dynamical systems
Keywords:
Irregular Time Series Forecasting, Temporal Patching

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_W5mDOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper argues that irregular-time-series patching combines two distinct choices: a support kernel determining which observations enter a patch and an aggregation measure determining their mass. Existing methods often learn the support while giving each record equal weight, thereby estimating an observation-intensity-weighted average rather than a physical-time average. The paper expresses this discrepancy as the patch-local covariance between signal content and observation intensity divided by mean intensity. VQPatch addresses it using boundary-clipped Voronoi cell widths as quadrature weights, separates content from the observation process, and later reintroduces sampling information through bounded conditioning. A second result gives an O(h) approximation rate under regularity assumptions. Across four benchmarks and 17 baselines, VQPatch achieves the best MAE on every dataset and the best result on seven of eight dataset-metric pairs. Controlled studies connect the theoretical covariance term to measured bias and show improved robustness under a shifted observation schedule.

Critical strengths
  1. The paper clearly separates which observations enter a patch from how much weight they receive. The covariance identity explains how sampling patterns can change the resulting average.
  2. Voronoi weights offer a simple correction without extra trainable parameters. The separate OPC pathway also preserves information about when measurements were taken.
  3. The evidence includes 17 baselines, five-seed main results, component tests, and a controlled change in sampling times. The model reports the lowest MAE on all four datasets.
Critical weaknesses
  1. A better estimate of a physical-time average does not necessarily give a better forecast. OPC preserves sampling information, but the paper needs stronger evidence about when this separation helps, especially when frequent measurements signal important events.
  2. The text overstates what the covariance identity proves. Zero covariance does not require independent recording, and the identity alone does not show that added model capacity can never compensate for the difference. Those broader claims need qualification.
  3. The sampling-shift test uses one dataset and one seed, while component tests also use one seed. Some reported gains are small, so repeated, paired comparisons would help establish how reliable they are.
Questions for authors

See Weakness

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

The paper proposes VQPatch for irregular time-series forecasting. It argues that conventional patch aggregation gives equal weight to observations regardless of their temporal coverage, introducing bias when sampling density varies. VQPatch uses Voronoi cell widths to weight observations and incorporates sampling information through a separate conditioning module. The paper provides a theoretical analysis of aggregation bias and evaluates the method against 17 baselines on four datasets.

Critical strengths

The paper identifies an interesting issue in irregular time-series patching: equal weighting of observations does not necessarily reflect their coverage in physical time. The proposed Voronoi weighting is simple and has a clear theoretical motivation. The experiments are extensive, with strong results across four datasets and competitive efficiency. The controlled sampling-shift experiments are also useful.

Critical weaknesses
  • The paper argues that aggregation bias cannot be corrected by increasing model capacity. However, Proposition 1 characterizes the difference between two aggregation measures, rather than proving that downstream models cannot compensate for it. This claim seems too strong.

  • The approximation guarantee in Proposition 2 relies on smoothness assumptions and a sufficiently small maximum Voronoi cell width. For highly sparse observations, a single measurement may represent a long interval. The practical behavior under such conditions deserves further analysis.

  • The ablation study in Table 2 uses only one random seed, and some differences are small. For example, removing VTQ increases PhysioNet MSE from 0.313920 to 0.314552. Multi-seed ablations would provide stronger evidence for the individual components.

  • The controlled schedule-shift experiments in Appendix D are limited to HumanActivity and one seed. Since the motivation also concerns informative clinical sampling, similar tests on clinical datasets would strengthen the conclusions.

  • The method assumes that physical-time weighting is preferable for content representation. However, densely sampled intervals may contain important rapid changes. The paper does not fully examine whether downweighting these observations can hurt forecasting, particularly for extreme values.

Questions for authors
  • Figure 3 shows that intermediate quadrature powers can outperform the default gamma = 1. Does this suggest that exact physical-time weighting is not always the best choice for forecasting?

  • How does VQPatch perform on sequences with very large observation gaps? Can the authors report results grouped by observation sparsity or maximum Voronoi cell width?

  • Would the improvement from VTQ remain significant if the ablations in Table 2 were repeated over five seeds?

  • Can the controlled sampling-shift results be extended to MIMIC or PhysioNet?

Reviewer_8cMTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

In the context of patching for Irregular Multivariate Time Series (IMTS) forecasting, existing approaches (such as t-PatchGNN and APN) have primarily focused on refining the temporal scope or "support" of the patches. Against this backdrop, the authors demonstrate that the discrepancy between "point-normalized aggregation" and "physical-time averaging" is equivalent to the covariance between the local patch content and the observation intensity, divided by the mean intensity. This bias is deterministic and it cannot be eliminated simply by increasing model capacity or the volume of training data. Consequently, the authors decompose patch aggregation into two distinct components: a "Support Kernel" and an "Aggregation Measure," and provide a theoretical analysis of the bias introduced by observation density. Based on this analysis, the paper proposes VQPatch, which assigns integration weights to each observation point according to the temporal span of its corresponding 1D Voronoi cell. Additionally, it incorporates information such as observation frequency and time intervals as auxiliary inputs to the forecasting model. The methodology is evaluated on four irregular time series datasets: HumanActivity, USHCN, PhysioNet, and MIMIC against 17 baseline models. The study also includes ablation studies and experiments examining the impact of varying sampling distributions.

Critical strengths

1- By decomposing the patch operator into "support" and "measure" components, the paper highlights that recent research has focused exclusively on optimizing the former while neglecting the latter. Rather than merely adding new network modules, the authors address the issue at the level of input representation as a valuable perspective.

2- The method is both simple and effective. Voronoi weights are derived solely from timestamps without introducing any learnable parameters. The additional computational overhead is linear, O(BS)O(BS), and the complexity analysis (Table 8) confirms that it does not increase the overall computational order. The main model requires only 0.002M–0.03M parameters and 0.4–0.66 GB of VRAM two orders of magnitude smaller than ASTGI or t-PatchGNN. For instance, on the PhysioNet dataset, VQPatch reduces peak training VRAM usage by approximately 61.4% compared to APN.

3- The observed performance gains align well with the theoretical predictions. The authors characterize the deviation of equal-weight aggregation from physical time-averaging by utilizing the covariance between observed intensity and temporal content, and subsequently employ Voronoi cell widths to construct corresponding numerical integration weights.

Critical weaknesses

1- While the adaptation of Voronoi tessellation/nearest-neighbor quadrature from spatial sampling to a one-dimensional time axis is elegantly executed, the concept of weighting by the width of a representative cell is a classic approach in numerical analysis and point process theory. It remains insufficiently explained whether weighting by physical time is necessarily optimal for prediction. Furthermore, there is a lack of direct comparison with simpler temporal weighting methods.

2- The effect sizes observed in the ablation studies are relatively small, and all experiments were conducted using a single random seed. Relying on a fixed, single seed makes it impossible to distinguish between genuine effects and fluctuations stemming from the training process. Ablation studies should be performed using multiple random seeds, or alternatively, paired significance tests should be conducted.

3- Voronoi weighting is essentially a form of nearest-neighbor rectangular quadrature, where the error scales with the maximum cell width. The authors acknowledge that this has not been quantified; it is recommended that the distribution of Voronoi cell sizes for each dataset be included in the main text, allowing readers to assess the extent to which this error bound holds.

4- The study relies on only four datasets with small sample sizes. The claim that the framework can be generalized to tasks such as classification, imputation, and event modeling lacks any experimental support. It is recommended that this claim be either removed or reclassified as "future work."

Questions for authors

Refer to the weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.